xref: /linux/net/ceph/osd_client.c (revision 507e190946297c34a27d9366b0661d5e506fdd03)
1 
2 #include <linux/ceph/ceph_debug.h>
3 
4 #include <linux/module.h>
5 #include <linux/err.h>
6 #include <linux/highmem.h>
7 #include <linux/mm.h>
8 #include <linux/pagemap.h>
9 #include <linux/slab.h>
10 #include <linux/uaccess.h>
11 #ifdef CONFIG_BLOCK
12 #include <linux/bio.h>
13 #endif
14 
15 #include <linux/ceph/ceph_features.h>
16 #include <linux/ceph/libceph.h>
17 #include <linux/ceph/osd_client.h>
18 #include <linux/ceph/messenger.h>
19 #include <linux/ceph/decode.h>
20 #include <linux/ceph/auth.h>
21 #include <linux/ceph/pagelist.h>
22 
23 #define OSD_OPREPLY_FRONT_LEN	512
24 
25 static struct kmem_cache	*ceph_osd_request_cache;
26 
27 static const struct ceph_connection_operations osd_con_ops;
28 
29 /*
30  * Implement client access to distributed object storage cluster.
31  *
32  * All data objects are stored within a cluster/cloud of OSDs, or
33  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
34  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
35  * remote daemons serving up and coordinating consistent and safe
36  * access to storage.
37  *
38  * Cluster membership and the mapping of data objects onto storage devices
39  * are described by the osd map.
40  *
41  * We keep track of pending OSD requests (read, write), resubmit
42  * requests to different OSDs when the cluster topology/data layout
43  * change, or retry the affected requests when the communications
44  * channel with an OSD is reset.
45  */
46 
47 static void link_request(struct ceph_osd *osd, struct ceph_osd_request *req);
48 static void unlink_request(struct ceph_osd *osd, struct ceph_osd_request *req);
49 static void link_linger(struct ceph_osd *osd,
50 			struct ceph_osd_linger_request *lreq);
51 static void unlink_linger(struct ceph_osd *osd,
52 			  struct ceph_osd_linger_request *lreq);
53 static void clear_backoffs(struct ceph_osd *osd);
54 
55 #if 1
56 static inline bool rwsem_is_wrlocked(struct rw_semaphore *sem)
57 {
58 	bool wrlocked = true;
59 
60 	if (unlikely(down_read_trylock(sem))) {
61 		wrlocked = false;
62 		up_read(sem);
63 	}
64 
65 	return wrlocked;
66 }
67 static inline void verify_osdc_locked(struct ceph_osd_client *osdc)
68 {
69 	WARN_ON(!rwsem_is_locked(&osdc->lock));
70 }
71 static inline void verify_osdc_wrlocked(struct ceph_osd_client *osdc)
72 {
73 	WARN_ON(!rwsem_is_wrlocked(&osdc->lock));
74 }
75 static inline void verify_osd_locked(struct ceph_osd *osd)
76 {
77 	struct ceph_osd_client *osdc = osd->o_osdc;
78 
79 	WARN_ON(!(mutex_is_locked(&osd->lock) &&
80 		  rwsem_is_locked(&osdc->lock)) &&
81 		!rwsem_is_wrlocked(&osdc->lock));
82 }
83 static inline void verify_lreq_locked(struct ceph_osd_linger_request *lreq)
84 {
85 	WARN_ON(!mutex_is_locked(&lreq->lock));
86 }
87 #else
88 static inline void verify_osdc_locked(struct ceph_osd_client *osdc) { }
89 static inline void verify_osdc_wrlocked(struct ceph_osd_client *osdc) { }
90 static inline void verify_osd_locked(struct ceph_osd *osd) { }
91 static inline void verify_lreq_locked(struct ceph_osd_linger_request *lreq) { }
92 #endif
93 
94 /*
95  * calculate the mapping of a file extent onto an object, and fill out the
96  * request accordingly.  shorten extent as necessary if it crosses an
97  * object boundary.
98  *
99  * fill osd op in request message.
100  */
101 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
102 			u64 *objnum, u64 *objoff, u64 *objlen)
103 {
104 	u64 orig_len = *plen;
105 	int r;
106 
107 	/* object extent? */
108 	r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
109 					  objoff, objlen);
110 	if (r < 0)
111 		return r;
112 	if (*objlen < orig_len) {
113 		*plen = *objlen;
114 		dout(" skipping last %llu, final file extent %llu~%llu\n",
115 		     orig_len - *plen, off, *plen);
116 	}
117 
118 	dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
119 
120 	return 0;
121 }
122 
123 static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
124 {
125 	memset(osd_data, 0, sizeof (*osd_data));
126 	osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
127 }
128 
129 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
130 			struct page **pages, u64 length, u32 alignment,
131 			bool pages_from_pool, bool own_pages)
132 {
133 	osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
134 	osd_data->pages = pages;
135 	osd_data->length = length;
136 	osd_data->alignment = alignment;
137 	osd_data->pages_from_pool = pages_from_pool;
138 	osd_data->own_pages = own_pages;
139 }
140 
141 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
142 			struct ceph_pagelist *pagelist)
143 {
144 	osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
145 	osd_data->pagelist = pagelist;
146 }
147 
148 #ifdef CONFIG_BLOCK
149 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
150 			struct bio *bio, size_t bio_length)
151 {
152 	osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
153 	osd_data->bio = bio;
154 	osd_data->bio_length = bio_length;
155 }
156 #endif /* CONFIG_BLOCK */
157 
158 #define osd_req_op_data(oreq, whch, typ, fld)				\
159 ({									\
160 	struct ceph_osd_request *__oreq = (oreq);			\
161 	unsigned int __whch = (whch);					\
162 	BUG_ON(__whch >= __oreq->r_num_ops);				\
163 	&__oreq->r_ops[__whch].typ.fld;					\
164 })
165 
166 static struct ceph_osd_data *
167 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
168 {
169 	BUG_ON(which >= osd_req->r_num_ops);
170 
171 	return &osd_req->r_ops[which].raw_data_in;
172 }
173 
174 struct ceph_osd_data *
175 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
176 			unsigned int which)
177 {
178 	return osd_req_op_data(osd_req, which, extent, osd_data);
179 }
180 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
181 
182 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
183 			unsigned int which, struct page **pages,
184 			u64 length, u32 alignment,
185 			bool pages_from_pool, bool own_pages)
186 {
187 	struct ceph_osd_data *osd_data;
188 
189 	osd_data = osd_req_op_raw_data_in(osd_req, which);
190 	ceph_osd_data_pages_init(osd_data, pages, length, alignment,
191 				pages_from_pool, own_pages);
192 }
193 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
194 
195 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
196 			unsigned int which, struct page **pages,
197 			u64 length, u32 alignment,
198 			bool pages_from_pool, bool own_pages)
199 {
200 	struct ceph_osd_data *osd_data;
201 
202 	osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
203 	ceph_osd_data_pages_init(osd_data, pages, length, alignment,
204 				pages_from_pool, own_pages);
205 }
206 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
207 
208 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
209 			unsigned int which, struct ceph_pagelist *pagelist)
210 {
211 	struct ceph_osd_data *osd_data;
212 
213 	osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
214 	ceph_osd_data_pagelist_init(osd_data, pagelist);
215 }
216 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
217 
218 #ifdef CONFIG_BLOCK
219 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
220 			unsigned int which, struct bio *bio, size_t bio_length)
221 {
222 	struct ceph_osd_data *osd_data;
223 
224 	osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
225 	ceph_osd_data_bio_init(osd_data, bio, bio_length);
226 }
227 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
228 #endif /* CONFIG_BLOCK */
229 
230 static void osd_req_op_cls_request_info_pagelist(
231 			struct ceph_osd_request *osd_req,
232 			unsigned int which, struct ceph_pagelist *pagelist)
233 {
234 	struct ceph_osd_data *osd_data;
235 
236 	osd_data = osd_req_op_data(osd_req, which, cls, request_info);
237 	ceph_osd_data_pagelist_init(osd_data, pagelist);
238 }
239 
240 void osd_req_op_cls_request_data_pagelist(
241 			struct ceph_osd_request *osd_req,
242 			unsigned int which, struct ceph_pagelist *pagelist)
243 {
244 	struct ceph_osd_data *osd_data;
245 
246 	osd_data = osd_req_op_data(osd_req, which, cls, request_data);
247 	ceph_osd_data_pagelist_init(osd_data, pagelist);
248 	osd_req->r_ops[which].cls.indata_len += pagelist->length;
249 	osd_req->r_ops[which].indata_len += pagelist->length;
250 }
251 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
252 
253 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
254 			unsigned int which, struct page **pages, u64 length,
255 			u32 alignment, bool pages_from_pool, bool own_pages)
256 {
257 	struct ceph_osd_data *osd_data;
258 
259 	osd_data = osd_req_op_data(osd_req, which, cls, request_data);
260 	ceph_osd_data_pages_init(osd_data, pages, length, alignment,
261 				pages_from_pool, own_pages);
262 	osd_req->r_ops[which].cls.indata_len += length;
263 	osd_req->r_ops[which].indata_len += length;
264 }
265 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
266 
267 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
268 			unsigned int which, struct page **pages, u64 length,
269 			u32 alignment, bool pages_from_pool, bool own_pages)
270 {
271 	struct ceph_osd_data *osd_data;
272 
273 	osd_data = osd_req_op_data(osd_req, which, cls, response_data);
274 	ceph_osd_data_pages_init(osd_data, pages, length, alignment,
275 				pages_from_pool, own_pages);
276 }
277 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
278 
279 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
280 {
281 	switch (osd_data->type) {
282 	case CEPH_OSD_DATA_TYPE_NONE:
283 		return 0;
284 	case CEPH_OSD_DATA_TYPE_PAGES:
285 		return osd_data->length;
286 	case CEPH_OSD_DATA_TYPE_PAGELIST:
287 		return (u64)osd_data->pagelist->length;
288 #ifdef CONFIG_BLOCK
289 	case CEPH_OSD_DATA_TYPE_BIO:
290 		return (u64)osd_data->bio_length;
291 #endif /* CONFIG_BLOCK */
292 	default:
293 		WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
294 		return 0;
295 	}
296 }
297 
298 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
299 {
300 	if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
301 		int num_pages;
302 
303 		num_pages = calc_pages_for((u64)osd_data->alignment,
304 						(u64)osd_data->length);
305 		ceph_release_page_vector(osd_data->pages, num_pages);
306 	}
307 	ceph_osd_data_init(osd_data);
308 }
309 
310 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
311 			unsigned int which)
312 {
313 	struct ceph_osd_req_op *op;
314 
315 	BUG_ON(which >= osd_req->r_num_ops);
316 	op = &osd_req->r_ops[which];
317 
318 	switch (op->op) {
319 	case CEPH_OSD_OP_READ:
320 	case CEPH_OSD_OP_WRITE:
321 	case CEPH_OSD_OP_WRITEFULL:
322 		ceph_osd_data_release(&op->extent.osd_data);
323 		break;
324 	case CEPH_OSD_OP_CALL:
325 		ceph_osd_data_release(&op->cls.request_info);
326 		ceph_osd_data_release(&op->cls.request_data);
327 		ceph_osd_data_release(&op->cls.response_data);
328 		break;
329 	case CEPH_OSD_OP_SETXATTR:
330 	case CEPH_OSD_OP_CMPXATTR:
331 		ceph_osd_data_release(&op->xattr.osd_data);
332 		break;
333 	case CEPH_OSD_OP_STAT:
334 		ceph_osd_data_release(&op->raw_data_in);
335 		break;
336 	case CEPH_OSD_OP_NOTIFY_ACK:
337 		ceph_osd_data_release(&op->notify_ack.request_data);
338 		break;
339 	case CEPH_OSD_OP_NOTIFY:
340 		ceph_osd_data_release(&op->notify.request_data);
341 		ceph_osd_data_release(&op->notify.response_data);
342 		break;
343 	case CEPH_OSD_OP_LIST_WATCHERS:
344 		ceph_osd_data_release(&op->list_watchers.response_data);
345 		break;
346 	default:
347 		break;
348 	}
349 }
350 
351 /*
352  * Assumes @t is zero-initialized.
353  */
354 static void target_init(struct ceph_osd_request_target *t)
355 {
356 	ceph_oid_init(&t->base_oid);
357 	ceph_oloc_init(&t->base_oloc);
358 	ceph_oid_init(&t->target_oid);
359 	ceph_oloc_init(&t->target_oloc);
360 
361 	ceph_osds_init(&t->acting);
362 	ceph_osds_init(&t->up);
363 	t->size = -1;
364 	t->min_size = -1;
365 
366 	t->osd = CEPH_HOMELESS_OSD;
367 }
368 
369 static void target_copy(struct ceph_osd_request_target *dest,
370 			const struct ceph_osd_request_target *src)
371 {
372 	ceph_oid_copy(&dest->base_oid, &src->base_oid);
373 	ceph_oloc_copy(&dest->base_oloc, &src->base_oloc);
374 	ceph_oid_copy(&dest->target_oid, &src->target_oid);
375 	ceph_oloc_copy(&dest->target_oloc, &src->target_oloc);
376 
377 	dest->pgid = src->pgid; /* struct */
378 	dest->spgid = src->spgid; /* struct */
379 	dest->pg_num = src->pg_num;
380 	dest->pg_num_mask = src->pg_num_mask;
381 	ceph_osds_copy(&dest->acting, &src->acting);
382 	ceph_osds_copy(&dest->up, &src->up);
383 	dest->size = src->size;
384 	dest->min_size = src->min_size;
385 	dest->sort_bitwise = src->sort_bitwise;
386 
387 	dest->flags = src->flags;
388 	dest->paused = src->paused;
389 
390 	dest->epoch = src->epoch;
391 	dest->last_force_resend = src->last_force_resend;
392 
393 	dest->osd = src->osd;
394 }
395 
396 static void target_destroy(struct ceph_osd_request_target *t)
397 {
398 	ceph_oid_destroy(&t->base_oid);
399 	ceph_oloc_destroy(&t->base_oloc);
400 	ceph_oid_destroy(&t->target_oid);
401 	ceph_oloc_destroy(&t->target_oloc);
402 }
403 
404 /*
405  * requests
406  */
407 static void request_release_checks(struct ceph_osd_request *req)
408 {
409 	WARN_ON(!RB_EMPTY_NODE(&req->r_node));
410 	WARN_ON(!RB_EMPTY_NODE(&req->r_mc_node));
411 	WARN_ON(!list_empty(&req->r_unsafe_item));
412 	WARN_ON(req->r_osd);
413 }
414 
415 static void ceph_osdc_release_request(struct kref *kref)
416 {
417 	struct ceph_osd_request *req = container_of(kref,
418 					    struct ceph_osd_request, r_kref);
419 	unsigned int which;
420 
421 	dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
422 	     req->r_request, req->r_reply);
423 	request_release_checks(req);
424 
425 	if (req->r_request)
426 		ceph_msg_put(req->r_request);
427 	if (req->r_reply)
428 		ceph_msg_put(req->r_reply);
429 
430 	for (which = 0; which < req->r_num_ops; which++)
431 		osd_req_op_data_release(req, which);
432 
433 	target_destroy(&req->r_t);
434 	ceph_put_snap_context(req->r_snapc);
435 
436 	if (req->r_mempool)
437 		mempool_free(req, req->r_osdc->req_mempool);
438 	else if (req->r_num_ops <= CEPH_OSD_SLAB_OPS)
439 		kmem_cache_free(ceph_osd_request_cache, req);
440 	else
441 		kfree(req);
442 }
443 
444 void ceph_osdc_get_request(struct ceph_osd_request *req)
445 {
446 	dout("%s %p (was %d)\n", __func__, req,
447 	     kref_read(&req->r_kref));
448 	kref_get(&req->r_kref);
449 }
450 EXPORT_SYMBOL(ceph_osdc_get_request);
451 
452 void ceph_osdc_put_request(struct ceph_osd_request *req)
453 {
454 	if (req) {
455 		dout("%s %p (was %d)\n", __func__, req,
456 		     kref_read(&req->r_kref));
457 		kref_put(&req->r_kref, ceph_osdc_release_request);
458 	}
459 }
460 EXPORT_SYMBOL(ceph_osdc_put_request);
461 
462 static void request_init(struct ceph_osd_request *req)
463 {
464 	/* req only, each op is zeroed in _osd_req_op_init() */
465 	memset(req, 0, sizeof(*req));
466 
467 	kref_init(&req->r_kref);
468 	init_completion(&req->r_completion);
469 	RB_CLEAR_NODE(&req->r_node);
470 	RB_CLEAR_NODE(&req->r_mc_node);
471 	INIT_LIST_HEAD(&req->r_unsafe_item);
472 
473 	target_init(&req->r_t);
474 }
475 
476 /*
477  * This is ugly, but it allows us to reuse linger registration and ping
478  * requests, keeping the structure of the code around send_linger{_ping}()
479  * reasonable.  Setting up a min_nr=2 mempool for each linger request
480  * and dealing with copying ops (this blasts req only, watch op remains
481  * intact) isn't any better.
482  */
483 static void request_reinit(struct ceph_osd_request *req)
484 {
485 	struct ceph_osd_client *osdc = req->r_osdc;
486 	bool mempool = req->r_mempool;
487 	unsigned int num_ops = req->r_num_ops;
488 	u64 snapid = req->r_snapid;
489 	struct ceph_snap_context *snapc = req->r_snapc;
490 	bool linger = req->r_linger;
491 	struct ceph_msg *request_msg = req->r_request;
492 	struct ceph_msg *reply_msg = req->r_reply;
493 
494 	dout("%s req %p\n", __func__, req);
495 	WARN_ON(kref_read(&req->r_kref) != 1);
496 	request_release_checks(req);
497 
498 	WARN_ON(kref_read(&request_msg->kref) != 1);
499 	WARN_ON(kref_read(&reply_msg->kref) != 1);
500 	target_destroy(&req->r_t);
501 
502 	request_init(req);
503 	req->r_osdc = osdc;
504 	req->r_mempool = mempool;
505 	req->r_num_ops = num_ops;
506 	req->r_snapid = snapid;
507 	req->r_snapc = snapc;
508 	req->r_linger = linger;
509 	req->r_request = request_msg;
510 	req->r_reply = reply_msg;
511 }
512 
513 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
514 					       struct ceph_snap_context *snapc,
515 					       unsigned int num_ops,
516 					       bool use_mempool,
517 					       gfp_t gfp_flags)
518 {
519 	struct ceph_osd_request *req;
520 
521 	if (use_mempool) {
522 		BUG_ON(num_ops > CEPH_OSD_SLAB_OPS);
523 		req = mempool_alloc(osdc->req_mempool, gfp_flags);
524 	} else if (num_ops <= CEPH_OSD_SLAB_OPS) {
525 		req = kmem_cache_alloc(ceph_osd_request_cache, gfp_flags);
526 	} else {
527 		BUG_ON(num_ops > CEPH_OSD_MAX_OPS);
528 		req = kmalloc(sizeof(*req) + num_ops * sizeof(req->r_ops[0]),
529 			      gfp_flags);
530 	}
531 	if (unlikely(!req))
532 		return NULL;
533 
534 	request_init(req);
535 	req->r_osdc = osdc;
536 	req->r_mempool = use_mempool;
537 	req->r_num_ops = num_ops;
538 	req->r_snapid = CEPH_NOSNAP;
539 	req->r_snapc = ceph_get_snap_context(snapc);
540 
541 	dout("%s req %p\n", __func__, req);
542 	return req;
543 }
544 EXPORT_SYMBOL(ceph_osdc_alloc_request);
545 
546 static int ceph_oloc_encoding_size(const struct ceph_object_locator *oloc)
547 {
548 	return 8 + 4 + 4 + 4 + (oloc->pool_ns ? oloc->pool_ns->len : 0);
549 }
550 
551 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp)
552 {
553 	struct ceph_osd_client *osdc = req->r_osdc;
554 	struct ceph_msg *msg;
555 	int msg_size;
556 
557 	WARN_ON(ceph_oid_empty(&req->r_base_oid));
558 	WARN_ON(ceph_oloc_empty(&req->r_base_oloc));
559 
560 	/* create request message */
561 	msg_size = CEPH_ENCODING_START_BLK_LEN +
562 			CEPH_PGID_ENCODING_LEN + 1; /* spgid */
563 	msg_size += 4 + 4 + 4; /* hash, osdmap_epoch, flags */
564 	msg_size += CEPH_ENCODING_START_BLK_LEN +
565 			sizeof(struct ceph_osd_reqid); /* reqid */
566 	msg_size += sizeof(struct ceph_blkin_trace_info); /* trace */
567 	msg_size += 4 + sizeof(struct ceph_timespec); /* client_inc, mtime */
568 	msg_size += CEPH_ENCODING_START_BLK_LEN +
569 			ceph_oloc_encoding_size(&req->r_base_oloc); /* oloc */
570 	msg_size += 4 + req->r_base_oid.name_len; /* oid */
571 	msg_size += 2 + req->r_num_ops * sizeof(struct ceph_osd_op);
572 	msg_size += 8; /* snapid */
573 	msg_size += 8; /* snap_seq */
574 	msg_size += 4 + 8 * (req->r_snapc ? req->r_snapc->num_snaps : 0);
575 	msg_size += 4 + 8; /* retry_attempt, features */
576 
577 	if (req->r_mempool)
578 		msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
579 	else
580 		msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp, true);
581 	if (!msg)
582 		return -ENOMEM;
583 
584 	memset(msg->front.iov_base, 0, msg->front.iov_len);
585 	req->r_request = msg;
586 
587 	/* create reply message */
588 	msg_size = OSD_OPREPLY_FRONT_LEN;
589 	msg_size += req->r_base_oid.name_len;
590 	msg_size += req->r_num_ops * sizeof(struct ceph_osd_op);
591 
592 	if (req->r_mempool)
593 		msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
594 	else
595 		msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, msg_size, gfp, true);
596 	if (!msg)
597 		return -ENOMEM;
598 
599 	req->r_reply = msg;
600 
601 	return 0;
602 }
603 EXPORT_SYMBOL(ceph_osdc_alloc_messages);
604 
605 static bool osd_req_opcode_valid(u16 opcode)
606 {
607 	switch (opcode) {
608 #define GENERATE_CASE(op, opcode, str)	case CEPH_OSD_OP_##op: return true;
609 __CEPH_FORALL_OSD_OPS(GENERATE_CASE)
610 #undef GENERATE_CASE
611 	default:
612 		return false;
613 	}
614 }
615 
616 /*
617  * This is an osd op init function for opcodes that have no data or
618  * other information associated with them.  It also serves as a
619  * common init routine for all the other init functions, below.
620  */
621 static struct ceph_osd_req_op *
622 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
623 		 u16 opcode, u32 flags)
624 {
625 	struct ceph_osd_req_op *op;
626 
627 	BUG_ON(which >= osd_req->r_num_ops);
628 	BUG_ON(!osd_req_opcode_valid(opcode));
629 
630 	op = &osd_req->r_ops[which];
631 	memset(op, 0, sizeof (*op));
632 	op->op = opcode;
633 	op->flags = flags;
634 
635 	return op;
636 }
637 
638 void osd_req_op_init(struct ceph_osd_request *osd_req,
639 		     unsigned int which, u16 opcode, u32 flags)
640 {
641 	(void)_osd_req_op_init(osd_req, which, opcode, flags);
642 }
643 EXPORT_SYMBOL(osd_req_op_init);
644 
645 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
646 				unsigned int which, u16 opcode,
647 				u64 offset, u64 length,
648 				u64 truncate_size, u32 truncate_seq)
649 {
650 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
651 						      opcode, 0);
652 	size_t payload_len = 0;
653 
654 	BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
655 	       opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO &&
656 	       opcode != CEPH_OSD_OP_TRUNCATE);
657 
658 	op->extent.offset = offset;
659 	op->extent.length = length;
660 	op->extent.truncate_size = truncate_size;
661 	op->extent.truncate_seq = truncate_seq;
662 	if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL)
663 		payload_len += length;
664 
665 	op->indata_len = payload_len;
666 }
667 EXPORT_SYMBOL(osd_req_op_extent_init);
668 
669 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
670 				unsigned int which, u64 length)
671 {
672 	struct ceph_osd_req_op *op;
673 	u64 previous;
674 
675 	BUG_ON(which >= osd_req->r_num_ops);
676 	op = &osd_req->r_ops[which];
677 	previous = op->extent.length;
678 
679 	if (length == previous)
680 		return;		/* Nothing to do */
681 	BUG_ON(length > previous);
682 
683 	op->extent.length = length;
684 	if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL)
685 		op->indata_len -= previous - length;
686 }
687 EXPORT_SYMBOL(osd_req_op_extent_update);
688 
689 void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
690 				unsigned int which, u64 offset_inc)
691 {
692 	struct ceph_osd_req_op *op, *prev_op;
693 
694 	BUG_ON(which + 1 >= osd_req->r_num_ops);
695 
696 	prev_op = &osd_req->r_ops[which];
697 	op = _osd_req_op_init(osd_req, which + 1, prev_op->op, prev_op->flags);
698 	/* dup previous one */
699 	op->indata_len = prev_op->indata_len;
700 	op->outdata_len = prev_op->outdata_len;
701 	op->extent = prev_op->extent;
702 	/* adjust offset */
703 	op->extent.offset += offset_inc;
704 	op->extent.length -= offset_inc;
705 
706 	if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL)
707 		op->indata_len -= offset_inc;
708 }
709 EXPORT_SYMBOL(osd_req_op_extent_dup_last);
710 
711 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
712 			u16 opcode, const char *class, const char *method)
713 {
714 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
715 						      opcode, 0);
716 	struct ceph_pagelist *pagelist;
717 	size_t payload_len = 0;
718 	size_t size;
719 
720 	BUG_ON(opcode != CEPH_OSD_OP_CALL);
721 
722 	pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
723 	BUG_ON(!pagelist);
724 	ceph_pagelist_init(pagelist);
725 
726 	op->cls.class_name = class;
727 	size = strlen(class);
728 	BUG_ON(size > (size_t) U8_MAX);
729 	op->cls.class_len = size;
730 	ceph_pagelist_append(pagelist, class, size);
731 	payload_len += size;
732 
733 	op->cls.method_name = method;
734 	size = strlen(method);
735 	BUG_ON(size > (size_t) U8_MAX);
736 	op->cls.method_len = size;
737 	ceph_pagelist_append(pagelist, method, size);
738 	payload_len += size;
739 
740 	osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
741 
742 	op->indata_len = payload_len;
743 }
744 EXPORT_SYMBOL(osd_req_op_cls_init);
745 
746 int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
747 			  u16 opcode, const char *name, const void *value,
748 			  size_t size, u8 cmp_op, u8 cmp_mode)
749 {
750 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
751 						      opcode, 0);
752 	struct ceph_pagelist *pagelist;
753 	size_t payload_len;
754 
755 	BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR);
756 
757 	pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
758 	if (!pagelist)
759 		return -ENOMEM;
760 
761 	ceph_pagelist_init(pagelist);
762 
763 	payload_len = strlen(name);
764 	op->xattr.name_len = payload_len;
765 	ceph_pagelist_append(pagelist, name, payload_len);
766 
767 	op->xattr.value_len = size;
768 	ceph_pagelist_append(pagelist, value, size);
769 	payload_len += size;
770 
771 	op->xattr.cmp_op = cmp_op;
772 	op->xattr.cmp_mode = cmp_mode;
773 
774 	ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist);
775 	op->indata_len = payload_len;
776 	return 0;
777 }
778 EXPORT_SYMBOL(osd_req_op_xattr_init);
779 
780 /*
781  * @watch_opcode: CEPH_OSD_WATCH_OP_*
782  */
783 static void osd_req_op_watch_init(struct ceph_osd_request *req, int which,
784 				  u64 cookie, u8 watch_opcode)
785 {
786 	struct ceph_osd_req_op *op;
787 
788 	op = _osd_req_op_init(req, which, CEPH_OSD_OP_WATCH, 0);
789 	op->watch.cookie = cookie;
790 	op->watch.op = watch_opcode;
791 	op->watch.gen = 0;
792 }
793 
794 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
795 				unsigned int which,
796 				u64 expected_object_size,
797 				u64 expected_write_size)
798 {
799 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
800 						      CEPH_OSD_OP_SETALLOCHINT,
801 						      0);
802 
803 	op->alloc_hint.expected_object_size = expected_object_size;
804 	op->alloc_hint.expected_write_size = expected_write_size;
805 
806 	/*
807 	 * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
808 	 * not worth a feature bit.  Set FAILOK per-op flag to make
809 	 * sure older osds don't trip over an unsupported opcode.
810 	 */
811 	op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
812 }
813 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
814 
815 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
816 				struct ceph_osd_data *osd_data)
817 {
818 	u64 length = ceph_osd_data_length(osd_data);
819 
820 	if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
821 		BUG_ON(length > (u64) SIZE_MAX);
822 		if (length)
823 			ceph_msg_data_add_pages(msg, osd_data->pages,
824 					length, osd_data->alignment);
825 	} else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
826 		BUG_ON(!length);
827 		ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
828 #ifdef CONFIG_BLOCK
829 	} else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
830 		ceph_msg_data_add_bio(msg, osd_data->bio, length);
831 #endif
832 	} else {
833 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
834 	}
835 }
836 
837 static u32 osd_req_encode_op(struct ceph_osd_op *dst,
838 			     const struct ceph_osd_req_op *src)
839 {
840 	if (WARN_ON(!osd_req_opcode_valid(src->op))) {
841 		pr_err("unrecognized osd opcode %d\n", src->op);
842 
843 		return 0;
844 	}
845 
846 	switch (src->op) {
847 	case CEPH_OSD_OP_STAT:
848 		break;
849 	case CEPH_OSD_OP_READ:
850 	case CEPH_OSD_OP_WRITE:
851 	case CEPH_OSD_OP_WRITEFULL:
852 	case CEPH_OSD_OP_ZERO:
853 	case CEPH_OSD_OP_TRUNCATE:
854 		dst->extent.offset = cpu_to_le64(src->extent.offset);
855 		dst->extent.length = cpu_to_le64(src->extent.length);
856 		dst->extent.truncate_size =
857 			cpu_to_le64(src->extent.truncate_size);
858 		dst->extent.truncate_seq =
859 			cpu_to_le32(src->extent.truncate_seq);
860 		break;
861 	case CEPH_OSD_OP_CALL:
862 		dst->cls.class_len = src->cls.class_len;
863 		dst->cls.method_len = src->cls.method_len;
864 		dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
865 		break;
866 	case CEPH_OSD_OP_STARTSYNC:
867 		break;
868 	case CEPH_OSD_OP_WATCH:
869 		dst->watch.cookie = cpu_to_le64(src->watch.cookie);
870 		dst->watch.ver = cpu_to_le64(0);
871 		dst->watch.op = src->watch.op;
872 		dst->watch.gen = cpu_to_le32(src->watch.gen);
873 		break;
874 	case CEPH_OSD_OP_NOTIFY_ACK:
875 		break;
876 	case CEPH_OSD_OP_NOTIFY:
877 		dst->notify.cookie = cpu_to_le64(src->notify.cookie);
878 		break;
879 	case CEPH_OSD_OP_LIST_WATCHERS:
880 		break;
881 	case CEPH_OSD_OP_SETALLOCHINT:
882 		dst->alloc_hint.expected_object_size =
883 		    cpu_to_le64(src->alloc_hint.expected_object_size);
884 		dst->alloc_hint.expected_write_size =
885 		    cpu_to_le64(src->alloc_hint.expected_write_size);
886 		break;
887 	case CEPH_OSD_OP_SETXATTR:
888 	case CEPH_OSD_OP_CMPXATTR:
889 		dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
890 		dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
891 		dst->xattr.cmp_op = src->xattr.cmp_op;
892 		dst->xattr.cmp_mode = src->xattr.cmp_mode;
893 		break;
894 	case CEPH_OSD_OP_CREATE:
895 	case CEPH_OSD_OP_DELETE:
896 		break;
897 	default:
898 		pr_err("unsupported osd opcode %s\n",
899 			ceph_osd_op_name(src->op));
900 		WARN_ON(1);
901 
902 		return 0;
903 	}
904 
905 	dst->op = cpu_to_le16(src->op);
906 	dst->flags = cpu_to_le32(src->flags);
907 	dst->payload_len = cpu_to_le32(src->indata_len);
908 
909 	return src->indata_len;
910 }
911 
912 /*
913  * build new request AND message, calculate layout, and adjust file
914  * extent as needed.
915  *
916  * if the file was recently truncated, we include information about its
917  * old and new size so that the object can be updated appropriately.  (we
918  * avoid synchronously deleting truncated objects because it's slow.)
919  *
920  * if @do_sync, include a 'startsync' command so that the osd will flush
921  * data quickly.
922  */
923 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
924 					       struct ceph_file_layout *layout,
925 					       struct ceph_vino vino,
926 					       u64 off, u64 *plen,
927 					       unsigned int which, int num_ops,
928 					       int opcode, int flags,
929 					       struct ceph_snap_context *snapc,
930 					       u32 truncate_seq,
931 					       u64 truncate_size,
932 					       bool use_mempool)
933 {
934 	struct ceph_osd_request *req;
935 	u64 objnum = 0;
936 	u64 objoff = 0;
937 	u64 objlen = 0;
938 	int r;
939 
940 	BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
941 	       opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE &&
942 	       opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE);
943 
944 	req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
945 					GFP_NOFS);
946 	if (!req) {
947 		r = -ENOMEM;
948 		goto fail;
949 	}
950 
951 	/* calculate max write size */
952 	r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
953 	if (r)
954 		goto fail;
955 
956 	if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) {
957 		osd_req_op_init(req, which, opcode, 0);
958 	} else {
959 		u32 object_size = layout->object_size;
960 		u32 object_base = off - objoff;
961 		if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
962 			if (truncate_size <= object_base) {
963 				truncate_size = 0;
964 			} else {
965 				truncate_size -= object_base;
966 				if (truncate_size > object_size)
967 					truncate_size = object_size;
968 			}
969 		}
970 		osd_req_op_extent_init(req, which, opcode, objoff, objlen,
971 				       truncate_size, truncate_seq);
972 	}
973 
974 	req->r_abort_on_full = true;
975 	req->r_flags = flags;
976 	req->r_base_oloc.pool = layout->pool_id;
977 	req->r_base_oloc.pool_ns = ceph_try_get_string(layout->pool_ns);
978 	ceph_oid_printf(&req->r_base_oid, "%llx.%08llx", vino.ino, objnum);
979 
980 	req->r_snapid = vino.snap;
981 	if (flags & CEPH_OSD_FLAG_WRITE)
982 		req->r_data_offset = off;
983 
984 	r = ceph_osdc_alloc_messages(req, GFP_NOFS);
985 	if (r)
986 		goto fail;
987 
988 	return req;
989 
990 fail:
991 	ceph_osdc_put_request(req);
992 	return ERR_PTR(r);
993 }
994 EXPORT_SYMBOL(ceph_osdc_new_request);
995 
996 /*
997  * We keep osd requests in an rbtree, sorted by ->r_tid.
998  */
999 DEFINE_RB_FUNCS(request, struct ceph_osd_request, r_tid, r_node)
1000 DEFINE_RB_FUNCS(request_mc, struct ceph_osd_request, r_tid, r_mc_node)
1001 
1002 static bool osd_homeless(struct ceph_osd *osd)
1003 {
1004 	return osd->o_osd == CEPH_HOMELESS_OSD;
1005 }
1006 
1007 static bool osd_registered(struct ceph_osd *osd)
1008 {
1009 	verify_osdc_locked(osd->o_osdc);
1010 
1011 	return !RB_EMPTY_NODE(&osd->o_node);
1012 }
1013 
1014 /*
1015  * Assumes @osd is zero-initialized.
1016  */
1017 static void osd_init(struct ceph_osd *osd)
1018 {
1019 	refcount_set(&osd->o_ref, 1);
1020 	RB_CLEAR_NODE(&osd->o_node);
1021 	osd->o_requests = RB_ROOT;
1022 	osd->o_linger_requests = RB_ROOT;
1023 	osd->o_backoff_mappings = RB_ROOT;
1024 	osd->o_backoffs_by_id = RB_ROOT;
1025 	INIT_LIST_HEAD(&osd->o_osd_lru);
1026 	INIT_LIST_HEAD(&osd->o_keepalive_item);
1027 	osd->o_incarnation = 1;
1028 	mutex_init(&osd->lock);
1029 }
1030 
1031 static void osd_cleanup(struct ceph_osd *osd)
1032 {
1033 	WARN_ON(!RB_EMPTY_NODE(&osd->o_node));
1034 	WARN_ON(!RB_EMPTY_ROOT(&osd->o_requests));
1035 	WARN_ON(!RB_EMPTY_ROOT(&osd->o_linger_requests));
1036 	WARN_ON(!RB_EMPTY_ROOT(&osd->o_backoff_mappings));
1037 	WARN_ON(!RB_EMPTY_ROOT(&osd->o_backoffs_by_id));
1038 	WARN_ON(!list_empty(&osd->o_osd_lru));
1039 	WARN_ON(!list_empty(&osd->o_keepalive_item));
1040 
1041 	if (osd->o_auth.authorizer) {
1042 		WARN_ON(osd_homeless(osd));
1043 		ceph_auth_destroy_authorizer(osd->o_auth.authorizer);
1044 	}
1045 }
1046 
1047 /*
1048  * Track open sessions with osds.
1049  */
1050 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
1051 {
1052 	struct ceph_osd *osd;
1053 
1054 	WARN_ON(onum == CEPH_HOMELESS_OSD);
1055 
1056 	osd = kzalloc(sizeof(*osd), GFP_NOIO | __GFP_NOFAIL);
1057 	osd_init(osd);
1058 	osd->o_osdc = osdc;
1059 	osd->o_osd = onum;
1060 
1061 	ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1062 
1063 	return osd;
1064 }
1065 
1066 static struct ceph_osd *get_osd(struct ceph_osd *osd)
1067 {
1068 	if (refcount_inc_not_zero(&osd->o_ref)) {
1069 		dout("get_osd %p %d -> %d\n", osd, refcount_read(&osd->o_ref)-1,
1070 		     refcount_read(&osd->o_ref));
1071 		return osd;
1072 	} else {
1073 		dout("get_osd %p FAIL\n", osd);
1074 		return NULL;
1075 	}
1076 }
1077 
1078 static void put_osd(struct ceph_osd *osd)
1079 {
1080 	dout("put_osd %p %d -> %d\n", osd, refcount_read(&osd->o_ref),
1081 	     refcount_read(&osd->o_ref) - 1);
1082 	if (refcount_dec_and_test(&osd->o_ref)) {
1083 		osd_cleanup(osd);
1084 		kfree(osd);
1085 	}
1086 }
1087 
1088 DEFINE_RB_FUNCS(osd, struct ceph_osd, o_osd, o_node)
1089 
1090 static void __move_osd_to_lru(struct ceph_osd *osd)
1091 {
1092 	struct ceph_osd_client *osdc = osd->o_osdc;
1093 
1094 	dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd);
1095 	BUG_ON(!list_empty(&osd->o_osd_lru));
1096 
1097 	spin_lock(&osdc->osd_lru_lock);
1098 	list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1099 	spin_unlock(&osdc->osd_lru_lock);
1100 
1101 	osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl;
1102 }
1103 
1104 static void maybe_move_osd_to_lru(struct ceph_osd *osd)
1105 {
1106 	if (RB_EMPTY_ROOT(&osd->o_requests) &&
1107 	    RB_EMPTY_ROOT(&osd->o_linger_requests))
1108 		__move_osd_to_lru(osd);
1109 }
1110 
1111 static void __remove_osd_from_lru(struct ceph_osd *osd)
1112 {
1113 	struct ceph_osd_client *osdc = osd->o_osdc;
1114 
1115 	dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd);
1116 
1117 	spin_lock(&osdc->osd_lru_lock);
1118 	if (!list_empty(&osd->o_osd_lru))
1119 		list_del_init(&osd->o_osd_lru);
1120 	spin_unlock(&osdc->osd_lru_lock);
1121 }
1122 
1123 /*
1124  * Close the connection and assign any leftover requests to the
1125  * homeless session.
1126  */
1127 static void close_osd(struct ceph_osd *osd)
1128 {
1129 	struct ceph_osd_client *osdc = osd->o_osdc;
1130 	struct rb_node *n;
1131 
1132 	verify_osdc_wrlocked(osdc);
1133 	dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd);
1134 
1135 	ceph_con_close(&osd->o_con);
1136 
1137 	for (n = rb_first(&osd->o_requests); n; ) {
1138 		struct ceph_osd_request *req =
1139 		    rb_entry(n, struct ceph_osd_request, r_node);
1140 
1141 		n = rb_next(n); /* unlink_request() */
1142 
1143 		dout(" reassigning req %p tid %llu\n", req, req->r_tid);
1144 		unlink_request(osd, req);
1145 		link_request(&osdc->homeless_osd, req);
1146 	}
1147 	for (n = rb_first(&osd->o_linger_requests); n; ) {
1148 		struct ceph_osd_linger_request *lreq =
1149 		    rb_entry(n, struct ceph_osd_linger_request, node);
1150 
1151 		n = rb_next(n); /* unlink_linger() */
1152 
1153 		dout(" reassigning lreq %p linger_id %llu\n", lreq,
1154 		     lreq->linger_id);
1155 		unlink_linger(osd, lreq);
1156 		link_linger(&osdc->homeless_osd, lreq);
1157 	}
1158 	clear_backoffs(osd);
1159 
1160 	__remove_osd_from_lru(osd);
1161 	erase_osd(&osdc->osds, osd);
1162 	put_osd(osd);
1163 }
1164 
1165 /*
1166  * reset osd connect
1167  */
1168 static int reopen_osd(struct ceph_osd *osd)
1169 {
1170 	struct ceph_entity_addr *peer_addr;
1171 
1172 	dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd);
1173 
1174 	if (RB_EMPTY_ROOT(&osd->o_requests) &&
1175 	    RB_EMPTY_ROOT(&osd->o_linger_requests)) {
1176 		close_osd(osd);
1177 		return -ENODEV;
1178 	}
1179 
1180 	peer_addr = &osd->o_osdc->osdmap->osd_addr[osd->o_osd];
1181 	if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1182 			!ceph_con_opened(&osd->o_con)) {
1183 		struct rb_node *n;
1184 
1185 		dout("osd addr hasn't changed and connection never opened, "
1186 		     "letting msgr retry\n");
1187 		/* touch each r_stamp for handle_timeout()'s benfit */
1188 		for (n = rb_first(&osd->o_requests); n; n = rb_next(n)) {
1189 			struct ceph_osd_request *req =
1190 			    rb_entry(n, struct ceph_osd_request, r_node);
1191 			req->r_stamp = jiffies;
1192 		}
1193 
1194 		return -EAGAIN;
1195 	}
1196 
1197 	ceph_con_close(&osd->o_con);
1198 	ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1199 	osd->o_incarnation++;
1200 
1201 	return 0;
1202 }
1203 
1204 static struct ceph_osd *lookup_create_osd(struct ceph_osd_client *osdc, int o,
1205 					  bool wrlocked)
1206 {
1207 	struct ceph_osd *osd;
1208 
1209 	if (wrlocked)
1210 		verify_osdc_wrlocked(osdc);
1211 	else
1212 		verify_osdc_locked(osdc);
1213 
1214 	if (o != CEPH_HOMELESS_OSD)
1215 		osd = lookup_osd(&osdc->osds, o);
1216 	else
1217 		osd = &osdc->homeless_osd;
1218 	if (!osd) {
1219 		if (!wrlocked)
1220 			return ERR_PTR(-EAGAIN);
1221 
1222 		osd = create_osd(osdc, o);
1223 		insert_osd(&osdc->osds, osd);
1224 		ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd,
1225 			      &osdc->osdmap->osd_addr[osd->o_osd]);
1226 	}
1227 
1228 	dout("%s osdc %p osd%d -> osd %p\n", __func__, osdc, o, osd);
1229 	return osd;
1230 }
1231 
1232 /*
1233  * Create request <-> OSD session relation.
1234  *
1235  * @req has to be assigned a tid, @osd may be homeless.
1236  */
1237 static void link_request(struct ceph_osd *osd, struct ceph_osd_request *req)
1238 {
1239 	verify_osd_locked(osd);
1240 	WARN_ON(!req->r_tid || req->r_osd);
1241 	dout("%s osd %p osd%d req %p tid %llu\n", __func__, osd, osd->o_osd,
1242 	     req, req->r_tid);
1243 
1244 	if (!osd_homeless(osd))
1245 		__remove_osd_from_lru(osd);
1246 	else
1247 		atomic_inc(&osd->o_osdc->num_homeless);
1248 
1249 	get_osd(osd);
1250 	insert_request(&osd->o_requests, req);
1251 	req->r_osd = osd;
1252 }
1253 
1254 static void unlink_request(struct ceph_osd *osd, struct ceph_osd_request *req)
1255 {
1256 	verify_osd_locked(osd);
1257 	WARN_ON(req->r_osd != osd);
1258 	dout("%s osd %p osd%d req %p tid %llu\n", __func__, osd, osd->o_osd,
1259 	     req, req->r_tid);
1260 
1261 	req->r_osd = NULL;
1262 	erase_request(&osd->o_requests, req);
1263 	put_osd(osd);
1264 
1265 	if (!osd_homeless(osd))
1266 		maybe_move_osd_to_lru(osd);
1267 	else
1268 		atomic_dec(&osd->o_osdc->num_homeless);
1269 }
1270 
1271 static bool __pool_full(struct ceph_pg_pool_info *pi)
1272 {
1273 	return pi->flags & CEPH_POOL_FLAG_FULL;
1274 }
1275 
1276 static bool have_pool_full(struct ceph_osd_client *osdc)
1277 {
1278 	struct rb_node *n;
1279 
1280 	for (n = rb_first(&osdc->osdmap->pg_pools); n; n = rb_next(n)) {
1281 		struct ceph_pg_pool_info *pi =
1282 		    rb_entry(n, struct ceph_pg_pool_info, node);
1283 
1284 		if (__pool_full(pi))
1285 			return true;
1286 	}
1287 
1288 	return false;
1289 }
1290 
1291 static bool pool_full(struct ceph_osd_client *osdc, s64 pool_id)
1292 {
1293 	struct ceph_pg_pool_info *pi;
1294 
1295 	pi = ceph_pg_pool_by_id(osdc->osdmap, pool_id);
1296 	if (!pi)
1297 		return false;
1298 
1299 	return __pool_full(pi);
1300 }
1301 
1302 /*
1303  * Returns whether a request should be blocked from being sent
1304  * based on the current osdmap and osd_client settings.
1305  */
1306 static bool target_should_be_paused(struct ceph_osd_client *osdc,
1307 				    const struct ceph_osd_request_target *t,
1308 				    struct ceph_pg_pool_info *pi)
1309 {
1310 	bool pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD);
1311 	bool pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) ||
1312 		       ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) ||
1313 		       __pool_full(pi);
1314 
1315 	WARN_ON(pi->id != t->target_oloc.pool);
1316 	return ((t->flags & CEPH_OSD_FLAG_READ) && pauserd) ||
1317 	       ((t->flags & CEPH_OSD_FLAG_WRITE) && pausewr) ||
1318 	       (osdc->osdmap->epoch < osdc->epoch_barrier);
1319 }
1320 
1321 enum calc_target_result {
1322 	CALC_TARGET_NO_ACTION = 0,
1323 	CALC_TARGET_NEED_RESEND,
1324 	CALC_TARGET_POOL_DNE,
1325 };
1326 
1327 static enum calc_target_result calc_target(struct ceph_osd_client *osdc,
1328 					   struct ceph_osd_request_target *t,
1329 					   struct ceph_connection *con,
1330 					   bool any_change)
1331 {
1332 	struct ceph_pg_pool_info *pi;
1333 	struct ceph_pg pgid, last_pgid;
1334 	struct ceph_osds up, acting;
1335 	bool force_resend = false;
1336 	bool unpaused = false;
1337 	bool legacy_change;
1338 	bool split = false;
1339 	bool sort_bitwise = ceph_osdmap_flag(osdc, CEPH_OSDMAP_SORTBITWISE);
1340 	bool recovery_deletes = ceph_osdmap_flag(osdc,
1341 						 CEPH_OSDMAP_RECOVERY_DELETES);
1342 	enum calc_target_result ct_res;
1343 	int ret;
1344 
1345 	t->epoch = osdc->osdmap->epoch;
1346 	pi = ceph_pg_pool_by_id(osdc->osdmap, t->base_oloc.pool);
1347 	if (!pi) {
1348 		t->osd = CEPH_HOMELESS_OSD;
1349 		ct_res = CALC_TARGET_POOL_DNE;
1350 		goto out;
1351 	}
1352 
1353 	if (osdc->osdmap->epoch == pi->last_force_request_resend) {
1354 		if (t->last_force_resend < pi->last_force_request_resend) {
1355 			t->last_force_resend = pi->last_force_request_resend;
1356 			force_resend = true;
1357 		} else if (t->last_force_resend == 0) {
1358 			force_resend = true;
1359 		}
1360 	}
1361 
1362 	/* apply tiering */
1363 	ceph_oid_copy(&t->target_oid, &t->base_oid);
1364 	ceph_oloc_copy(&t->target_oloc, &t->base_oloc);
1365 	if ((t->flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1366 		if (t->flags & CEPH_OSD_FLAG_READ && pi->read_tier >= 0)
1367 			t->target_oloc.pool = pi->read_tier;
1368 		if (t->flags & CEPH_OSD_FLAG_WRITE && pi->write_tier >= 0)
1369 			t->target_oloc.pool = pi->write_tier;
1370 
1371 		pi = ceph_pg_pool_by_id(osdc->osdmap, t->target_oloc.pool);
1372 		if (!pi) {
1373 			t->osd = CEPH_HOMELESS_OSD;
1374 			ct_res = CALC_TARGET_POOL_DNE;
1375 			goto out;
1376 		}
1377 	}
1378 
1379 	ret = __ceph_object_locator_to_pg(pi, &t->target_oid, &t->target_oloc,
1380 					  &pgid);
1381 	if (ret) {
1382 		WARN_ON(ret != -ENOENT);
1383 		t->osd = CEPH_HOMELESS_OSD;
1384 		ct_res = CALC_TARGET_POOL_DNE;
1385 		goto out;
1386 	}
1387 	last_pgid.pool = pgid.pool;
1388 	last_pgid.seed = ceph_stable_mod(pgid.seed, t->pg_num, t->pg_num_mask);
1389 
1390 	ceph_pg_to_up_acting_osds(osdc->osdmap, pi, &pgid, &up, &acting);
1391 	if (any_change &&
1392 	    ceph_is_new_interval(&t->acting,
1393 				 &acting,
1394 				 &t->up,
1395 				 &up,
1396 				 t->size,
1397 				 pi->size,
1398 				 t->min_size,
1399 				 pi->min_size,
1400 				 t->pg_num,
1401 				 pi->pg_num,
1402 				 t->sort_bitwise,
1403 				 sort_bitwise,
1404 				 t->recovery_deletes,
1405 				 recovery_deletes,
1406 				 &last_pgid))
1407 		force_resend = true;
1408 
1409 	if (t->paused && !target_should_be_paused(osdc, t, pi)) {
1410 		t->paused = false;
1411 		unpaused = true;
1412 	}
1413 	legacy_change = ceph_pg_compare(&t->pgid, &pgid) ||
1414 			ceph_osds_changed(&t->acting, &acting, any_change);
1415 	if (t->pg_num)
1416 		split = ceph_pg_is_split(&last_pgid, t->pg_num, pi->pg_num);
1417 
1418 	if (legacy_change || force_resend || split) {
1419 		t->pgid = pgid; /* struct */
1420 		ceph_pg_to_primary_shard(osdc->osdmap, pi, &pgid, &t->spgid);
1421 		ceph_osds_copy(&t->acting, &acting);
1422 		ceph_osds_copy(&t->up, &up);
1423 		t->size = pi->size;
1424 		t->min_size = pi->min_size;
1425 		t->pg_num = pi->pg_num;
1426 		t->pg_num_mask = pi->pg_num_mask;
1427 		t->sort_bitwise = sort_bitwise;
1428 		t->recovery_deletes = recovery_deletes;
1429 
1430 		t->osd = acting.primary;
1431 	}
1432 
1433 	if (unpaused || legacy_change || force_resend ||
1434 	    (split && con && CEPH_HAVE_FEATURE(con->peer_features,
1435 					       RESEND_ON_SPLIT)))
1436 		ct_res = CALC_TARGET_NEED_RESEND;
1437 	else
1438 		ct_res = CALC_TARGET_NO_ACTION;
1439 
1440 out:
1441 	dout("%s t %p -> ct_res %d osd %d\n", __func__, t, ct_res, t->osd);
1442 	return ct_res;
1443 }
1444 
1445 static struct ceph_spg_mapping *alloc_spg_mapping(void)
1446 {
1447 	struct ceph_spg_mapping *spg;
1448 
1449 	spg = kmalloc(sizeof(*spg), GFP_NOIO);
1450 	if (!spg)
1451 		return NULL;
1452 
1453 	RB_CLEAR_NODE(&spg->node);
1454 	spg->backoffs = RB_ROOT;
1455 	return spg;
1456 }
1457 
1458 static void free_spg_mapping(struct ceph_spg_mapping *spg)
1459 {
1460 	WARN_ON(!RB_EMPTY_NODE(&spg->node));
1461 	WARN_ON(!RB_EMPTY_ROOT(&spg->backoffs));
1462 
1463 	kfree(spg);
1464 }
1465 
1466 /*
1467  * rbtree of ceph_spg_mapping for handling map<spg_t, ...>, similar to
1468  * ceph_pg_mapping.  Used to track OSD backoffs -- a backoff [range] is
1469  * defined only within a specific spgid; it does not pass anything to
1470  * children on split, or to another primary.
1471  */
1472 DEFINE_RB_FUNCS2(spg_mapping, struct ceph_spg_mapping, spgid, ceph_spg_compare,
1473 		 RB_BYPTR, const struct ceph_spg *, node)
1474 
1475 static u64 hoid_get_bitwise_key(const struct ceph_hobject_id *hoid)
1476 {
1477 	return hoid->is_max ? 0x100000000ull : hoid->hash_reverse_bits;
1478 }
1479 
1480 static void hoid_get_effective_key(const struct ceph_hobject_id *hoid,
1481 				   void **pkey, size_t *pkey_len)
1482 {
1483 	if (hoid->key_len) {
1484 		*pkey = hoid->key;
1485 		*pkey_len = hoid->key_len;
1486 	} else {
1487 		*pkey = hoid->oid;
1488 		*pkey_len = hoid->oid_len;
1489 	}
1490 }
1491 
1492 static int compare_names(const void *name1, size_t name1_len,
1493 			 const void *name2, size_t name2_len)
1494 {
1495 	int ret;
1496 
1497 	ret = memcmp(name1, name2, min(name1_len, name2_len));
1498 	if (!ret) {
1499 		if (name1_len < name2_len)
1500 			ret = -1;
1501 		else if (name1_len > name2_len)
1502 			ret = 1;
1503 	}
1504 	return ret;
1505 }
1506 
1507 static int hoid_compare(const struct ceph_hobject_id *lhs,
1508 			const struct ceph_hobject_id *rhs)
1509 {
1510 	void *effective_key1, *effective_key2;
1511 	size_t effective_key1_len, effective_key2_len;
1512 	int ret;
1513 
1514 	if (lhs->is_max < rhs->is_max)
1515 		return -1;
1516 	if (lhs->is_max > rhs->is_max)
1517 		return 1;
1518 
1519 	if (lhs->pool < rhs->pool)
1520 		return -1;
1521 	if (lhs->pool > rhs->pool)
1522 		return 1;
1523 
1524 	if (hoid_get_bitwise_key(lhs) < hoid_get_bitwise_key(rhs))
1525 		return -1;
1526 	if (hoid_get_bitwise_key(lhs) > hoid_get_bitwise_key(rhs))
1527 		return 1;
1528 
1529 	ret = compare_names(lhs->nspace, lhs->nspace_len,
1530 			    rhs->nspace, rhs->nspace_len);
1531 	if (ret)
1532 		return ret;
1533 
1534 	hoid_get_effective_key(lhs, &effective_key1, &effective_key1_len);
1535 	hoid_get_effective_key(rhs, &effective_key2, &effective_key2_len);
1536 	ret = compare_names(effective_key1, effective_key1_len,
1537 			    effective_key2, effective_key2_len);
1538 	if (ret)
1539 		return ret;
1540 
1541 	ret = compare_names(lhs->oid, lhs->oid_len, rhs->oid, rhs->oid_len);
1542 	if (ret)
1543 		return ret;
1544 
1545 	if (lhs->snapid < rhs->snapid)
1546 		return -1;
1547 	if (lhs->snapid > rhs->snapid)
1548 		return 1;
1549 
1550 	return 0;
1551 }
1552 
1553 /*
1554  * For decoding ->begin and ->end of MOSDBackoff only -- no MIN/MAX
1555  * compat stuff here.
1556  *
1557  * Assumes @hoid is zero-initialized.
1558  */
1559 static int decode_hoid(void **p, void *end, struct ceph_hobject_id *hoid)
1560 {
1561 	u8 struct_v;
1562 	u32 struct_len;
1563 	int ret;
1564 
1565 	ret = ceph_start_decoding(p, end, 4, "hobject_t", &struct_v,
1566 				  &struct_len);
1567 	if (ret)
1568 		return ret;
1569 
1570 	if (struct_v < 4) {
1571 		pr_err("got struct_v %d < 4 of hobject_t\n", struct_v);
1572 		goto e_inval;
1573 	}
1574 
1575 	hoid->key = ceph_extract_encoded_string(p, end, &hoid->key_len,
1576 						GFP_NOIO);
1577 	if (IS_ERR(hoid->key)) {
1578 		ret = PTR_ERR(hoid->key);
1579 		hoid->key = NULL;
1580 		return ret;
1581 	}
1582 
1583 	hoid->oid = ceph_extract_encoded_string(p, end, &hoid->oid_len,
1584 						GFP_NOIO);
1585 	if (IS_ERR(hoid->oid)) {
1586 		ret = PTR_ERR(hoid->oid);
1587 		hoid->oid = NULL;
1588 		return ret;
1589 	}
1590 
1591 	ceph_decode_64_safe(p, end, hoid->snapid, e_inval);
1592 	ceph_decode_32_safe(p, end, hoid->hash, e_inval);
1593 	ceph_decode_8_safe(p, end, hoid->is_max, e_inval);
1594 
1595 	hoid->nspace = ceph_extract_encoded_string(p, end, &hoid->nspace_len,
1596 						   GFP_NOIO);
1597 	if (IS_ERR(hoid->nspace)) {
1598 		ret = PTR_ERR(hoid->nspace);
1599 		hoid->nspace = NULL;
1600 		return ret;
1601 	}
1602 
1603 	ceph_decode_64_safe(p, end, hoid->pool, e_inval);
1604 
1605 	ceph_hoid_build_hash_cache(hoid);
1606 	return 0;
1607 
1608 e_inval:
1609 	return -EINVAL;
1610 }
1611 
1612 static int hoid_encoding_size(const struct ceph_hobject_id *hoid)
1613 {
1614 	return 8 + 4 + 1 + 8 + /* snapid, hash, is_max, pool */
1615 	       4 + hoid->key_len + 4 + hoid->oid_len + 4 + hoid->nspace_len;
1616 }
1617 
1618 static void encode_hoid(void **p, void *end, const struct ceph_hobject_id *hoid)
1619 {
1620 	ceph_start_encoding(p, 4, 3, hoid_encoding_size(hoid));
1621 	ceph_encode_string(p, end, hoid->key, hoid->key_len);
1622 	ceph_encode_string(p, end, hoid->oid, hoid->oid_len);
1623 	ceph_encode_64(p, hoid->snapid);
1624 	ceph_encode_32(p, hoid->hash);
1625 	ceph_encode_8(p, hoid->is_max);
1626 	ceph_encode_string(p, end, hoid->nspace, hoid->nspace_len);
1627 	ceph_encode_64(p, hoid->pool);
1628 }
1629 
1630 static void free_hoid(struct ceph_hobject_id *hoid)
1631 {
1632 	if (hoid) {
1633 		kfree(hoid->key);
1634 		kfree(hoid->oid);
1635 		kfree(hoid->nspace);
1636 		kfree(hoid);
1637 	}
1638 }
1639 
1640 static struct ceph_osd_backoff *alloc_backoff(void)
1641 {
1642 	struct ceph_osd_backoff *backoff;
1643 
1644 	backoff = kzalloc(sizeof(*backoff), GFP_NOIO);
1645 	if (!backoff)
1646 		return NULL;
1647 
1648 	RB_CLEAR_NODE(&backoff->spg_node);
1649 	RB_CLEAR_NODE(&backoff->id_node);
1650 	return backoff;
1651 }
1652 
1653 static void free_backoff(struct ceph_osd_backoff *backoff)
1654 {
1655 	WARN_ON(!RB_EMPTY_NODE(&backoff->spg_node));
1656 	WARN_ON(!RB_EMPTY_NODE(&backoff->id_node));
1657 
1658 	free_hoid(backoff->begin);
1659 	free_hoid(backoff->end);
1660 	kfree(backoff);
1661 }
1662 
1663 /*
1664  * Within a specific spgid, backoffs are managed by ->begin hoid.
1665  */
1666 DEFINE_RB_INSDEL_FUNCS2(backoff, struct ceph_osd_backoff, begin, hoid_compare,
1667 			RB_BYVAL, spg_node);
1668 
1669 static struct ceph_osd_backoff *lookup_containing_backoff(struct rb_root *root,
1670 					    const struct ceph_hobject_id *hoid)
1671 {
1672 	struct rb_node *n = root->rb_node;
1673 
1674 	while (n) {
1675 		struct ceph_osd_backoff *cur =
1676 		    rb_entry(n, struct ceph_osd_backoff, spg_node);
1677 		int cmp;
1678 
1679 		cmp = hoid_compare(hoid, cur->begin);
1680 		if (cmp < 0) {
1681 			n = n->rb_left;
1682 		} else if (cmp > 0) {
1683 			if (hoid_compare(hoid, cur->end) < 0)
1684 				return cur;
1685 
1686 			n = n->rb_right;
1687 		} else {
1688 			return cur;
1689 		}
1690 	}
1691 
1692 	return NULL;
1693 }
1694 
1695 /*
1696  * Each backoff has a unique id within its OSD session.
1697  */
1698 DEFINE_RB_FUNCS(backoff_by_id, struct ceph_osd_backoff, id, id_node)
1699 
1700 static void clear_backoffs(struct ceph_osd *osd)
1701 {
1702 	while (!RB_EMPTY_ROOT(&osd->o_backoff_mappings)) {
1703 		struct ceph_spg_mapping *spg =
1704 		    rb_entry(rb_first(&osd->o_backoff_mappings),
1705 			     struct ceph_spg_mapping, node);
1706 
1707 		while (!RB_EMPTY_ROOT(&spg->backoffs)) {
1708 			struct ceph_osd_backoff *backoff =
1709 			    rb_entry(rb_first(&spg->backoffs),
1710 				     struct ceph_osd_backoff, spg_node);
1711 
1712 			erase_backoff(&spg->backoffs, backoff);
1713 			erase_backoff_by_id(&osd->o_backoffs_by_id, backoff);
1714 			free_backoff(backoff);
1715 		}
1716 		erase_spg_mapping(&osd->o_backoff_mappings, spg);
1717 		free_spg_mapping(spg);
1718 	}
1719 }
1720 
1721 /*
1722  * Set up a temporary, non-owning view into @t.
1723  */
1724 static void hoid_fill_from_target(struct ceph_hobject_id *hoid,
1725 				  const struct ceph_osd_request_target *t)
1726 {
1727 	hoid->key = NULL;
1728 	hoid->key_len = 0;
1729 	hoid->oid = t->target_oid.name;
1730 	hoid->oid_len = t->target_oid.name_len;
1731 	hoid->snapid = CEPH_NOSNAP;
1732 	hoid->hash = t->pgid.seed;
1733 	hoid->is_max = false;
1734 	if (t->target_oloc.pool_ns) {
1735 		hoid->nspace = t->target_oloc.pool_ns->str;
1736 		hoid->nspace_len = t->target_oloc.pool_ns->len;
1737 	} else {
1738 		hoid->nspace = NULL;
1739 		hoid->nspace_len = 0;
1740 	}
1741 	hoid->pool = t->target_oloc.pool;
1742 	ceph_hoid_build_hash_cache(hoid);
1743 }
1744 
1745 static bool should_plug_request(struct ceph_osd_request *req)
1746 {
1747 	struct ceph_osd *osd = req->r_osd;
1748 	struct ceph_spg_mapping *spg;
1749 	struct ceph_osd_backoff *backoff;
1750 	struct ceph_hobject_id hoid;
1751 
1752 	spg = lookup_spg_mapping(&osd->o_backoff_mappings, &req->r_t.spgid);
1753 	if (!spg)
1754 		return false;
1755 
1756 	hoid_fill_from_target(&hoid, &req->r_t);
1757 	backoff = lookup_containing_backoff(&spg->backoffs, &hoid);
1758 	if (!backoff)
1759 		return false;
1760 
1761 	dout("%s req %p tid %llu backoff osd%d spgid %llu.%xs%d id %llu\n",
1762 	     __func__, req, req->r_tid, osd->o_osd, backoff->spgid.pgid.pool,
1763 	     backoff->spgid.pgid.seed, backoff->spgid.shard, backoff->id);
1764 	return true;
1765 }
1766 
1767 static void setup_request_data(struct ceph_osd_request *req,
1768 			       struct ceph_msg *msg)
1769 {
1770 	u32 data_len = 0;
1771 	int i;
1772 
1773 	if (!list_empty(&msg->data))
1774 		return;
1775 
1776 	WARN_ON(msg->data_length);
1777 	for (i = 0; i < req->r_num_ops; i++) {
1778 		struct ceph_osd_req_op *op = &req->r_ops[i];
1779 
1780 		switch (op->op) {
1781 		/* request */
1782 		case CEPH_OSD_OP_WRITE:
1783 		case CEPH_OSD_OP_WRITEFULL:
1784 			WARN_ON(op->indata_len != op->extent.length);
1785 			ceph_osdc_msg_data_add(msg, &op->extent.osd_data);
1786 			break;
1787 		case CEPH_OSD_OP_SETXATTR:
1788 		case CEPH_OSD_OP_CMPXATTR:
1789 			WARN_ON(op->indata_len != op->xattr.name_len +
1790 						  op->xattr.value_len);
1791 			ceph_osdc_msg_data_add(msg, &op->xattr.osd_data);
1792 			break;
1793 		case CEPH_OSD_OP_NOTIFY_ACK:
1794 			ceph_osdc_msg_data_add(msg,
1795 					       &op->notify_ack.request_data);
1796 			break;
1797 
1798 		/* reply */
1799 		case CEPH_OSD_OP_STAT:
1800 			ceph_osdc_msg_data_add(req->r_reply,
1801 					       &op->raw_data_in);
1802 			break;
1803 		case CEPH_OSD_OP_READ:
1804 			ceph_osdc_msg_data_add(req->r_reply,
1805 					       &op->extent.osd_data);
1806 			break;
1807 		case CEPH_OSD_OP_LIST_WATCHERS:
1808 			ceph_osdc_msg_data_add(req->r_reply,
1809 					       &op->list_watchers.response_data);
1810 			break;
1811 
1812 		/* both */
1813 		case CEPH_OSD_OP_CALL:
1814 			WARN_ON(op->indata_len != op->cls.class_len +
1815 						  op->cls.method_len +
1816 						  op->cls.indata_len);
1817 			ceph_osdc_msg_data_add(msg, &op->cls.request_info);
1818 			/* optional, can be NONE */
1819 			ceph_osdc_msg_data_add(msg, &op->cls.request_data);
1820 			/* optional, can be NONE */
1821 			ceph_osdc_msg_data_add(req->r_reply,
1822 					       &op->cls.response_data);
1823 			break;
1824 		case CEPH_OSD_OP_NOTIFY:
1825 			ceph_osdc_msg_data_add(msg,
1826 					       &op->notify.request_data);
1827 			ceph_osdc_msg_data_add(req->r_reply,
1828 					       &op->notify.response_data);
1829 			break;
1830 		}
1831 
1832 		data_len += op->indata_len;
1833 	}
1834 
1835 	WARN_ON(data_len != msg->data_length);
1836 }
1837 
1838 static void encode_pgid(void **p, const struct ceph_pg *pgid)
1839 {
1840 	ceph_encode_8(p, 1);
1841 	ceph_encode_64(p, pgid->pool);
1842 	ceph_encode_32(p, pgid->seed);
1843 	ceph_encode_32(p, -1); /* preferred */
1844 }
1845 
1846 static void encode_spgid(void **p, const struct ceph_spg *spgid)
1847 {
1848 	ceph_start_encoding(p, 1, 1, CEPH_PGID_ENCODING_LEN + 1);
1849 	encode_pgid(p, &spgid->pgid);
1850 	ceph_encode_8(p, spgid->shard);
1851 }
1852 
1853 static void encode_oloc(void **p, void *end,
1854 			const struct ceph_object_locator *oloc)
1855 {
1856 	ceph_start_encoding(p, 5, 4, ceph_oloc_encoding_size(oloc));
1857 	ceph_encode_64(p, oloc->pool);
1858 	ceph_encode_32(p, -1); /* preferred */
1859 	ceph_encode_32(p, 0);  /* key len */
1860 	if (oloc->pool_ns)
1861 		ceph_encode_string(p, end, oloc->pool_ns->str,
1862 				   oloc->pool_ns->len);
1863 	else
1864 		ceph_encode_32(p, 0);
1865 }
1866 
1867 static void encode_request_partial(struct ceph_osd_request *req,
1868 				   struct ceph_msg *msg)
1869 {
1870 	void *p = msg->front.iov_base;
1871 	void *const end = p + msg->front_alloc_len;
1872 	u32 data_len = 0;
1873 	int i;
1874 
1875 	if (req->r_flags & CEPH_OSD_FLAG_WRITE) {
1876 		/* snapshots aren't writeable */
1877 		WARN_ON(req->r_snapid != CEPH_NOSNAP);
1878 	} else {
1879 		WARN_ON(req->r_mtime.tv_sec || req->r_mtime.tv_nsec ||
1880 			req->r_data_offset || req->r_snapc);
1881 	}
1882 
1883 	setup_request_data(req, msg);
1884 
1885 	encode_spgid(&p, &req->r_t.spgid); /* actual spg */
1886 	ceph_encode_32(&p, req->r_t.pgid.seed); /* raw hash */
1887 	ceph_encode_32(&p, req->r_osdc->osdmap->epoch);
1888 	ceph_encode_32(&p, req->r_flags);
1889 
1890 	/* reqid */
1891 	ceph_start_encoding(&p, 2, 2, sizeof(struct ceph_osd_reqid));
1892 	memset(p, 0, sizeof(struct ceph_osd_reqid));
1893 	p += sizeof(struct ceph_osd_reqid);
1894 
1895 	/* trace */
1896 	memset(p, 0, sizeof(struct ceph_blkin_trace_info));
1897 	p += sizeof(struct ceph_blkin_trace_info);
1898 
1899 	ceph_encode_32(&p, 0); /* client_inc, always 0 */
1900 	ceph_encode_timespec(p, &req->r_mtime);
1901 	p += sizeof(struct ceph_timespec);
1902 
1903 	encode_oloc(&p, end, &req->r_t.target_oloc);
1904 	ceph_encode_string(&p, end, req->r_t.target_oid.name,
1905 			   req->r_t.target_oid.name_len);
1906 
1907 	/* ops, can imply data */
1908 	ceph_encode_16(&p, req->r_num_ops);
1909 	for (i = 0; i < req->r_num_ops; i++) {
1910 		data_len += osd_req_encode_op(p, &req->r_ops[i]);
1911 		p += sizeof(struct ceph_osd_op);
1912 	}
1913 
1914 	ceph_encode_64(&p, req->r_snapid); /* snapid */
1915 	if (req->r_snapc) {
1916 		ceph_encode_64(&p, req->r_snapc->seq);
1917 		ceph_encode_32(&p, req->r_snapc->num_snaps);
1918 		for (i = 0; i < req->r_snapc->num_snaps; i++)
1919 			ceph_encode_64(&p, req->r_snapc->snaps[i]);
1920 	} else {
1921 		ceph_encode_64(&p, 0); /* snap_seq */
1922 		ceph_encode_32(&p, 0); /* snaps len */
1923 	}
1924 
1925 	ceph_encode_32(&p, req->r_attempts); /* retry_attempt */
1926 	BUG_ON(p > end - 8); /* space for features */
1927 
1928 	msg->hdr.version = cpu_to_le16(8); /* MOSDOp v8 */
1929 	/* front_len is finalized in encode_request_finish() */
1930 	msg->front.iov_len = p - msg->front.iov_base;
1931 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1932 	msg->hdr.data_len = cpu_to_le32(data_len);
1933 	/*
1934 	 * The header "data_off" is a hint to the receiver allowing it
1935 	 * to align received data into its buffers such that there's no
1936 	 * need to re-copy it before writing it to disk (direct I/O).
1937 	 */
1938 	msg->hdr.data_off = cpu_to_le16(req->r_data_offset);
1939 
1940 	dout("%s req %p msg %p oid %s oid_len %d\n", __func__, req, msg,
1941 	     req->r_t.target_oid.name, req->r_t.target_oid.name_len);
1942 }
1943 
1944 static void encode_request_finish(struct ceph_msg *msg)
1945 {
1946 	void *p = msg->front.iov_base;
1947 	void *const partial_end = p + msg->front.iov_len;
1948 	void *const end = p + msg->front_alloc_len;
1949 
1950 	if (CEPH_HAVE_FEATURE(msg->con->peer_features, RESEND_ON_SPLIT)) {
1951 		/* luminous OSD -- encode features and be done */
1952 		p = partial_end;
1953 		ceph_encode_64(&p, msg->con->peer_features);
1954 	} else {
1955 		struct {
1956 			char spgid[CEPH_ENCODING_START_BLK_LEN +
1957 				   CEPH_PGID_ENCODING_LEN + 1];
1958 			__le32 hash;
1959 			__le32 epoch;
1960 			__le32 flags;
1961 			char reqid[CEPH_ENCODING_START_BLK_LEN +
1962 				   sizeof(struct ceph_osd_reqid)];
1963 			char trace[sizeof(struct ceph_blkin_trace_info)];
1964 			__le32 client_inc;
1965 			struct ceph_timespec mtime;
1966 		} __packed head;
1967 		struct ceph_pg pgid;
1968 		void *oloc, *oid, *tail;
1969 		int oloc_len, oid_len, tail_len;
1970 		int len;
1971 
1972 		/*
1973 		 * Pre-luminous OSD -- reencode v8 into v4 using @head
1974 		 * as a temporary buffer.  Encode the raw PG; the rest
1975 		 * is just a matter of moving oloc, oid and tail blobs
1976 		 * around.
1977 		 */
1978 		memcpy(&head, p, sizeof(head));
1979 		p += sizeof(head);
1980 
1981 		oloc = p;
1982 		p += CEPH_ENCODING_START_BLK_LEN;
1983 		pgid.pool = ceph_decode_64(&p);
1984 		p += 4 + 4; /* preferred, key len */
1985 		len = ceph_decode_32(&p);
1986 		p += len;   /* nspace */
1987 		oloc_len = p - oloc;
1988 
1989 		oid = p;
1990 		len = ceph_decode_32(&p);
1991 		p += len;
1992 		oid_len = p - oid;
1993 
1994 		tail = p;
1995 		tail_len = partial_end - p;
1996 
1997 		p = msg->front.iov_base;
1998 		ceph_encode_copy(&p, &head.client_inc, sizeof(head.client_inc));
1999 		ceph_encode_copy(&p, &head.epoch, sizeof(head.epoch));
2000 		ceph_encode_copy(&p, &head.flags, sizeof(head.flags));
2001 		ceph_encode_copy(&p, &head.mtime, sizeof(head.mtime));
2002 
2003 		/* reassert_version */
2004 		memset(p, 0, sizeof(struct ceph_eversion));
2005 		p += sizeof(struct ceph_eversion);
2006 
2007 		BUG_ON(p >= oloc);
2008 		memmove(p, oloc, oloc_len);
2009 		p += oloc_len;
2010 
2011 		pgid.seed = le32_to_cpu(head.hash);
2012 		encode_pgid(&p, &pgid); /* raw pg */
2013 
2014 		BUG_ON(p >= oid);
2015 		memmove(p, oid, oid_len);
2016 		p += oid_len;
2017 
2018 		/* tail -- ops, snapid, snapc, retry_attempt */
2019 		BUG_ON(p >= tail);
2020 		memmove(p, tail, tail_len);
2021 		p += tail_len;
2022 
2023 		msg->hdr.version = cpu_to_le16(4); /* MOSDOp v4 */
2024 	}
2025 
2026 	BUG_ON(p > end);
2027 	msg->front.iov_len = p - msg->front.iov_base;
2028 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2029 
2030 	dout("%s msg %p tid %llu %u+%u+%u v%d\n", __func__, msg,
2031 	     le64_to_cpu(msg->hdr.tid), le32_to_cpu(msg->hdr.front_len),
2032 	     le32_to_cpu(msg->hdr.middle_len), le32_to_cpu(msg->hdr.data_len),
2033 	     le16_to_cpu(msg->hdr.version));
2034 }
2035 
2036 /*
2037  * @req has to be assigned a tid and registered.
2038  */
2039 static void send_request(struct ceph_osd_request *req)
2040 {
2041 	struct ceph_osd *osd = req->r_osd;
2042 
2043 	verify_osd_locked(osd);
2044 	WARN_ON(osd->o_osd != req->r_t.osd);
2045 
2046 	/* backoff? */
2047 	if (should_plug_request(req))
2048 		return;
2049 
2050 	/*
2051 	 * We may have a previously queued request message hanging
2052 	 * around.  Cancel it to avoid corrupting the msgr.
2053 	 */
2054 	if (req->r_sent)
2055 		ceph_msg_revoke(req->r_request);
2056 
2057 	req->r_flags |= CEPH_OSD_FLAG_KNOWN_REDIR;
2058 	if (req->r_attempts)
2059 		req->r_flags |= CEPH_OSD_FLAG_RETRY;
2060 	else
2061 		WARN_ON(req->r_flags & CEPH_OSD_FLAG_RETRY);
2062 
2063 	encode_request_partial(req, req->r_request);
2064 
2065 	dout("%s req %p tid %llu to pgid %llu.%x spgid %llu.%xs%d osd%d e%u flags 0x%x attempt %d\n",
2066 	     __func__, req, req->r_tid, req->r_t.pgid.pool, req->r_t.pgid.seed,
2067 	     req->r_t.spgid.pgid.pool, req->r_t.spgid.pgid.seed,
2068 	     req->r_t.spgid.shard, osd->o_osd, req->r_t.epoch, req->r_flags,
2069 	     req->r_attempts);
2070 
2071 	req->r_t.paused = false;
2072 	req->r_stamp = jiffies;
2073 	req->r_attempts++;
2074 
2075 	req->r_sent = osd->o_incarnation;
2076 	req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
2077 	ceph_con_send(&osd->o_con, ceph_msg_get(req->r_request));
2078 }
2079 
2080 static void maybe_request_map(struct ceph_osd_client *osdc)
2081 {
2082 	bool continuous = false;
2083 
2084 	verify_osdc_locked(osdc);
2085 	WARN_ON(!osdc->osdmap->epoch);
2086 
2087 	if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) ||
2088 	    ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD) ||
2089 	    ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR)) {
2090 		dout("%s osdc %p continuous\n", __func__, osdc);
2091 		continuous = true;
2092 	} else {
2093 		dout("%s osdc %p onetime\n", __func__, osdc);
2094 	}
2095 
2096 	if (ceph_monc_want_map(&osdc->client->monc, CEPH_SUB_OSDMAP,
2097 			       osdc->osdmap->epoch + 1, continuous))
2098 		ceph_monc_renew_subs(&osdc->client->monc);
2099 }
2100 
2101 static void complete_request(struct ceph_osd_request *req, int err);
2102 static void send_map_check(struct ceph_osd_request *req);
2103 
2104 static void __submit_request(struct ceph_osd_request *req, bool wrlocked)
2105 {
2106 	struct ceph_osd_client *osdc = req->r_osdc;
2107 	struct ceph_osd *osd;
2108 	enum calc_target_result ct_res;
2109 	bool need_send = false;
2110 	bool promoted = false;
2111 	bool need_abort = false;
2112 
2113 	WARN_ON(req->r_tid);
2114 	dout("%s req %p wrlocked %d\n", __func__, req, wrlocked);
2115 
2116 again:
2117 	ct_res = calc_target(osdc, &req->r_t, NULL, false);
2118 	if (ct_res == CALC_TARGET_POOL_DNE && !wrlocked)
2119 		goto promote;
2120 
2121 	osd = lookup_create_osd(osdc, req->r_t.osd, wrlocked);
2122 	if (IS_ERR(osd)) {
2123 		WARN_ON(PTR_ERR(osd) != -EAGAIN || wrlocked);
2124 		goto promote;
2125 	}
2126 
2127 	if (osdc->osdmap->epoch < osdc->epoch_barrier) {
2128 		dout("req %p epoch %u barrier %u\n", req, osdc->osdmap->epoch,
2129 		     osdc->epoch_barrier);
2130 		req->r_t.paused = true;
2131 		maybe_request_map(osdc);
2132 	} else if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
2133 		   ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR)) {
2134 		dout("req %p pausewr\n", req);
2135 		req->r_t.paused = true;
2136 		maybe_request_map(osdc);
2137 	} else if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
2138 		   ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD)) {
2139 		dout("req %p pauserd\n", req);
2140 		req->r_t.paused = true;
2141 		maybe_request_map(osdc);
2142 	} else if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
2143 		   !(req->r_flags & (CEPH_OSD_FLAG_FULL_TRY |
2144 				     CEPH_OSD_FLAG_FULL_FORCE)) &&
2145 		   (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) ||
2146 		    pool_full(osdc, req->r_t.base_oloc.pool))) {
2147 		dout("req %p full/pool_full\n", req);
2148 		pr_warn_ratelimited("FULL or reached pool quota\n");
2149 		req->r_t.paused = true;
2150 		maybe_request_map(osdc);
2151 		if (req->r_abort_on_full)
2152 			need_abort = true;
2153 	} else if (!osd_homeless(osd)) {
2154 		need_send = true;
2155 	} else {
2156 		maybe_request_map(osdc);
2157 	}
2158 
2159 	mutex_lock(&osd->lock);
2160 	/*
2161 	 * Assign the tid atomically with send_request() to protect
2162 	 * multiple writes to the same object from racing with each
2163 	 * other, resulting in out of order ops on the OSDs.
2164 	 */
2165 	req->r_tid = atomic64_inc_return(&osdc->last_tid);
2166 	link_request(osd, req);
2167 	if (need_send)
2168 		send_request(req);
2169 	else if (need_abort)
2170 		complete_request(req, -ENOSPC);
2171 	mutex_unlock(&osd->lock);
2172 
2173 	if (ct_res == CALC_TARGET_POOL_DNE)
2174 		send_map_check(req);
2175 
2176 	if (promoted)
2177 		downgrade_write(&osdc->lock);
2178 	return;
2179 
2180 promote:
2181 	up_read(&osdc->lock);
2182 	down_write(&osdc->lock);
2183 	wrlocked = true;
2184 	promoted = true;
2185 	goto again;
2186 }
2187 
2188 static void account_request(struct ceph_osd_request *req)
2189 {
2190 	WARN_ON(req->r_flags & (CEPH_OSD_FLAG_ACK | CEPH_OSD_FLAG_ONDISK));
2191 	WARN_ON(!(req->r_flags & (CEPH_OSD_FLAG_READ | CEPH_OSD_FLAG_WRITE)));
2192 
2193 	req->r_flags |= CEPH_OSD_FLAG_ONDISK;
2194 	atomic_inc(&req->r_osdc->num_requests);
2195 
2196 	req->r_start_stamp = jiffies;
2197 }
2198 
2199 static void submit_request(struct ceph_osd_request *req, bool wrlocked)
2200 {
2201 	ceph_osdc_get_request(req);
2202 	account_request(req);
2203 	__submit_request(req, wrlocked);
2204 }
2205 
2206 static void finish_request(struct ceph_osd_request *req)
2207 {
2208 	struct ceph_osd_client *osdc = req->r_osdc;
2209 
2210 	WARN_ON(lookup_request_mc(&osdc->map_checks, req->r_tid));
2211 	dout("%s req %p tid %llu\n", __func__, req, req->r_tid);
2212 
2213 	if (req->r_osd)
2214 		unlink_request(req->r_osd, req);
2215 	atomic_dec(&osdc->num_requests);
2216 
2217 	/*
2218 	 * If an OSD has failed or returned and a request has been sent
2219 	 * twice, it's possible to get a reply and end up here while the
2220 	 * request message is queued for delivery.  We will ignore the
2221 	 * reply, so not a big deal, but better to try and catch it.
2222 	 */
2223 	ceph_msg_revoke(req->r_request);
2224 	ceph_msg_revoke_incoming(req->r_reply);
2225 }
2226 
2227 static void __complete_request(struct ceph_osd_request *req)
2228 {
2229 	if (req->r_callback) {
2230 		dout("%s req %p tid %llu cb %pf result %d\n", __func__, req,
2231 		     req->r_tid, req->r_callback, req->r_result);
2232 		req->r_callback(req);
2233 	}
2234 }
2235 
2236 /*
2237  * This is open-coded in handle_reply().
2238  */
2239 static void complete_request(struct ceph_osd_request *req, int err)
2240 {
2241 	dout("%s req %p tid %llu err %d\n", __func__, req, req->r_tid, err);
2242 
2243 	req->r_result = err;
2244 	finish_request(req);
2245 	__complete_request(req);
2246 	complete_all(&req->r_completion);
2247 	ceph_osdc_put_request(req);
2248 }
2249 
2250 static void cancel_map_check(struct ceph_osd_request *req)
2251 {
2252 	struct ceph_osd_client *osdc = req->r_osdc;
2253 	struct ceph_osd_request *lookup_req;
2254 
2255 	verify_osdc_wrlocked(osdc);
2256 
2257 	lookup_req = lookup_request_mc(&osdc->map_checks, req->r_tid);
2258 	if (!lookup_req)
2259 		return;
2260 
2261 	WARN_ON(lookup_req != req);
2262 	erase_request_mc(&osdc->map_checks, req);
2263 	ceph_osdc_put_request(req);
2264 }
2265 
2266 static void cancel_request(struct ceph_osd_request *req)
2267 {
2268 	dout("%s req %p tid %llu\n", __func__, req, req->r_tid);
2269 
2270 	cancel_map_check(req);
2271 	finish_request(req);
2272 	complete_all(&req->r_completion);
2273 	ceph_osdc_put_request(req);
2274 }
2275 
2276 static void abort_request(struct ceph_osd_request *req, int err)
2277 {
2278 	dout("%s req %p tid %llu err %d\n", __func__, req, req->r_tid, err);
2279 
2280 	cancel_map_check(req);
2281 	complete_request(req, err);
2282 }
2283 
2284 static void update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb)
2285 {
2286 	if (likely(eb > osdc->epoch_barrier)) {
2287 		dout("updating epoch_barrier from %u to %u\n",
2288 				osdc->epoch_barrier, eb);
2289 		osdc->epoch_barrier = eb;
2290 		/* Request map if we're not to the barrier yet */
2291 		if (eb > osdc->osdmap->epoch)
2292 			maybe_request_map(osdc);
2293 	}
2294 }
2295 
2296 void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb)
2297 {
2298 	down_read(&osdc->lock);
2299 	if (unlikely(eb > osdc->epoch_barrier)) {
2300 		up_read(&osdc->lock);
2301 		down_write(&osdc->lock);
2302 		update_epoch_barrier(osdc, eb);
2303 		up_write(&osdc->lock);
2304 	} else {
2305 		up_read(&osdc->lock);
2306 	}
2307 }
2308 EXPORT_SYMBOL(ceph_osdc_update_epoch_barrier);
2309 
2310 /*
2311  * Drop all pending requests that are stalled waiting on a full condition to
2312  * clear, and complete them with ENOSPC as the return code. Set the
2313  * osdc->epoch_barrier to the latest map epoch that we've seen if any were
2314  * cancelled.
2315  */
2316 static void ceph_osdc_abort_on_full(struct ceph_osd_client *osdc)
2317 {
2318 	struct rb_node *n;
2319 	bool victims = false;
2320 
2321 	dout("enter abort_on_full\n");
2322 
2323 	if (!ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) && !have_pool_full(osdc))
2324 		goto out;
2325 
2326 	/* Scan list and see if there is anything to abort */
2327 	for (n = rb_first(&osdc->osds); n; n = rb_next(n)) {
2328 		struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node);
2329 		struct rb_node *m;
2330 
2331 		m = rb_first(&osd->o_requests);
2332 		while (m) {
2333 			struct ceph_osd_request *req = rb_entry(m,
2334 					struct ceph_osd_request, r_node);
2335 			m = rb_next(m);
2336 
2337 			if (req->r_abort_on_full) {
2338 				victims = true;
2339 				break;
2340 			}
2341 		}
2342 		if (victims)
2343 			break;
2344 	}
2345 
2346 	if (!victims)
2347 		goto out;
2348 
2349 	/*
2350 	 * Update the barrier to current epoch if it's behind that point,
2351 	 * since we know we have some calls to be aborted in the tree.
2352 	 */
2353 	update_epoch_barrier(osdc, osdc->osdmap->epoch);
2354 
2355 	for (n = rb_first(&osdc->osds); n; n = rb_next(n)) {
2356 		struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node);
2357 		struct rb_node *m;
2358 
2359 		m = rb_first(&osd->o_requests);
2360 		while (m) {
2361 			struct ceph_osd_request *req = rb_entry(m,
2362 					struct ceph_osd_request, r_node);
2363 			m = rb_next(m);
2364 
2365 			if (req->r_abort_on_full &&
2366 			    (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) ||
2367 			     pool_full(osdc, req->r_t.target_oloc.pool)))
2368 				abort_request(req, -ENOSPC);
2369 		}
2370 	}
2371 out:
2372 	dout("return abort_on_full barrier=%u\n", osdc->epoch_barrier);
2373 }
2374 
2375 static void check_pool_dne(struct ceph_osd_request *req)
2376 {
2377 	struct ceph_osd_client *osdc = req->r_osdc;
2378 	struct ceph_osdmap *map = osdc->osdmap;
2379 
2380 	verify_osdc_wrlocked(osdc);
2381 	WARN_ON(!map->epoch);
2382 
2383 	if (req->r_attempts) {
2384 		/*
2385 		 * We sent a request earlier, which means that
2386 		 * previously the pool existed, and now it does not
2387 		 * (i.e., it was deleted).
2388 		 */
2389 		req->r_map_dne_bound = map->epoch;
2390 		dout("%s req %p tid %llu pool disappeared\n", __func__, req,
2391 		     req->r_tid);
2392 	} else {
2393 		dout("%s req %p tid %llu map_dne_bound %u have %u\n", __func__,
2394 		     req, req->r_tid, req->r_map_dne_bound, map->epoch);
2395 	}
2396 
2397 	if (req->r_map_dne_bound) {
2398 		if (map->epoch >= req->r_map_dne_bound) {
2399 			/* we had a new enough map */
2400 			pr_info_ratelimited("tid %llu pool does not exist\n",
2401 					    req->r_tid);
2402 			complete_request(req, -ENOENT);
2403 		}
2404 	} else {
2405 		send_map_check(req);
2406 	}
2407 }
2408 
2409 static void map_check_cb(struct ceph_mon_generic_request *greq)
2410 {
2411 	struct ceph_osd_client *osdc = &greq->monc->client->osdc;
2412 	struct ceph_osd_request *req;
2413 	u64 tid = greq->private_data;
2414 
2415 	WARN_ON(greq->result || !greq->u.newest);
2416 
2417 	down_write(&osdc->lock);
2418 	req = lookup_request_mc(&osdc->map_checks, tid);
2419 	if (!req) {
2420 		dout("%s tid %llu dne\n", __func__, tid);
2421 		goto out_unlock;
2422 	}
2423 
2424 	dout("%s req %p tid %llu map_dne_bound %u newest %llu\n", __func__,
2425 	     req, req->r_tid, req->r_map_dne_bound, greq->u.newest);
2426 	if (!req->r_map_dne_bound)
2427 		req->r_map_dne_bound = greq->u.newest;
2428 	erase_request_mc(&osdc->map_checks, req);
2429 	check_pool_dne(req);
2430 
2431 	ceph_osdc_put_request(req);
2432 out_unlock:
2433 	up_write(&osdc->lock);
2434 }
2435 
2436 static void send_map_check(struct ceph_osd_request *req)
2437 {
2438 	struct ceph_osd_client *osdc = req->r_osdc;
2439 	struct ceph_osd_request *lookup_req;
2440 	int ret;
2441 
2442 	verify_osdc_wrlocked(osdc);
2443 
2444 	lookup_req = lookup_request_mc(&osdc->map_checks, req->r_tid);
2445 	if (lookup_req) {
2446 		WARN_ON(lookup_req != req);
2447 		return;
2448 	}
2449 
2450 	ceph_osdc_get_request(req);
2451 	insert_request_mc(&osdc->map_checks, req);
2452 	ret = ceph_monc_get_version_async(&osdc->client->monc, "osdmap",
2453 					  map_check_cb, req->r_tid);
2454 	WARN_ON(ret);
2455 }
2456 
2457 /*
2458  * lingering requests, watch/notify v2 infrastructure
2459  */
2460 static void linger_release(struct kref *kref)
2461 {
2462 	struct ceph_osd_linger_request *lreq =
2463 	    container_of(kref, struct ceph_osd_linger_request, kref);
2464 
2465 	dout("%s lreq %p reg_req %p ping_req %p\n", __func__, lreq,
2466 	     lreq->reg_req, lreq->ping_req);
2467 	WARN_ON(!RB_EMPTY_NODE(&lreq->node));
2468 	WARN_ON(!RB_EMPTY_NODE(&lreq->osdc_node));
2469 	WARN_ON(!RB_EMPTY_NODE(&lreq->mc_node));
2470 	WARN_ON(!list_empty(&lreq->scan_item));
2471 	WARN_ON(!list_empty(&lreq->pending_lworks));
2472 	WARN_ON(lreq->osd);
2473 
2474 	if (lreq->reg_req)
2475 		ceph_osdc_put_request(lreq->reg_req);
2476 	if (lreq->ping_req)
2477 		ceph_osdc_put_request(lreq->ping_req);
2478 	target_destroy(&lreq->t);
2479 	kfree(lreq);
2480 }
2481 
2482 static void linger_put(struct ceph_osd_linger_request *lreq)
2483 {
2484 	if (lreq)
2485 		kref_put(&lreq->kref, linger_release);
2486 }
2487 
2488 static struct ceph_osd_linger_request *
2489 linger_get(struct ceph_osd_linger_request *lreq)
2490 {
2491 	kref_get(&lreq->kref);
2492 	return lreq;
2493 }
2494 
2495 static struct ceph_osd_linger_request *
2496 linger_alloc(struct ceph_osd_client *osdc)
2497 {
2498 	struct ceph_osd_linger_request *lreq;
2499 
2500 	lreq = kzalloc(sizeof(*lreq), GFP_NOIO);
2501 	if (!lreq)
2502 		return NULL;
2503 
2504 	kref_init(&lreq->kref);
2505 	mutex_init(&lreq->lock);
2506 	RB_CLEAR_NODE(&lreq->node);
2507 	RB_CLEAR_NODE(&lreq->osdc_node);
2508 	RB_CLEAR_NODE(&lreq->mc_node);
2509 	INIT_LIST_HEAD(&lreq->scan_item);
2510 	INIT_LIST_HEAD(&lreq->pending_lworks);
2511 	init_completion(&lreq->reg_commit_wait);
2512 	init_completion(&lreq->notify_finish_wait);
2513 
2514 	lreq->osdc = osdc;
2515 	target_init(&lreq->t);
2516 
2517 	dout("%s lreq %p\n", __func__, lreq);
2518 	return lreq;
2519 }
2520 
2521 DEFINE_RB_INSDEL_FUNCS(linger, struct ceph_osd_linger_request, linger_id, node)
2522 DEFINE_RB_FUNCS(linger_osdc, struct ceph_osd_linger_request, linger_id, osdc_node)
2523 DEFINE_RB_FUNCS(linger_mc, struct ceph_osd_linger_request, linger_id, mc_node)
2524 
2525 /*
2526  * Create linger request <-> OSD session relation.
2527  *
2528  * @lreq has to be registered, @osd may be homeless.
2529  */
2530 static void link_linger(struct ceph_osd *osd,
2531 			struct ceph_osd_linger_request *lreq)
2532 {
2533 	verify_osd_locked(osd);
2534 	WARN_ON(!lreq->linger_id || lreq->osd);
2535 	dout("%s osd %p osd%d lreq %p linger_id %llu\n", __func__, osd,
2536 	     osd->o_osd, lreq, lreq->linger_id);
2537 
2538 	if (!osd_homeless(osd))
2539 		__remove_osd_from_lru(osd);
2540 	else
2541 		atomic_inc(&osd->o_osdc->num_homeless);
2542 
2543 	get_osd(osd);
2544 	insert_linger(&osd->o_linger_requests, lreq);
2545 	lreq->osd = osd;
2546 }
2547 
2548 static void unlink_linger(struct ceph_osd *osd,
2549 			  struct ceph_osd_linger_request *lreq)
2550 {
2551 	verify_osd_locked(osd);
2552 	WARN_ON(lreq->osd != osd);
2553 	dout("%s osd %p osd%d lreq %p linger_id %llu\n", __func__, osd,
2554 	     osd->o_osd, lreq, lreq->linger_id);
2555 
2556 	lreq->osd = NULL;
2557 	erase_linger(&osd->o_linger_requests, lreq);
2558 	put_osd(osd);
2559 
2560 	if (!osd_homeless(osd))
2561 		maybe_move_osd_to_lru(osd);
2562 	else
2563 		atomic_dec(&osd->o_osdc->num_homeless);
2564 }
2565 
2566 static bool __linger_registered(struct ceph_osd_linger_request *lreq)
2567 {
2568 	verify_osdc_locked(lreq->osdc);
2569 
2570 	return !RB_EMPTY_NODE(&lreq->osdc_node);
2571 }
2572 
2573 static bool linger_registered(struct ceph_osd_linger_request *lreq)
2574 {
2575 	struct ceph_osd_client *osdc = lreq->osdc;
2576 	bool registered;
2577 
2578 	down_read(&osdc->lock);
2579 	registered = __linger_registered(lreq);
2580 	up_read(&osdc->lock);
2581 
2582 	return registered;
2583 }
2584 
2585 static void linger_register(struct ceph_osd_linger_request *lreq)
2586 {
2587 	struct ceph_osd_client *osdc = lreq->osdc;
2588 
2589 	verify_osdc_wrlocked(osdc);
2590 	WARN_ON(lreq->linger_id);
2591 
2592 	linger_get(lreq);
2593 	lreq->linger_id = ++osdc->last_linger_id;
2594 	insert_linger_osdc(&osdc->linger_requests, lreq);
2595 }
2596 
2597 static void linger_unregister(struct ceph_osd_linger_request *lreq)
2598 {
2599 	struct ceph_osd_client *osdc = lreq->osdc;
2600 
2601 	verify_osdc_wrlocked(osdc);
2602 
2603 	erase_linger_osdc(&osdc->linger_requests, lreq);
2604 	linger_put(lreq);
2605 }
2606 
2607 static void cancel_linger_request(struct ceph_osd_request *req)
2608 {
2609 	struct ceph_osd_linger_request *lreq = req->r_priv;
2610 
2611 	WARN_ON(!req->r_linger);
2612 	cancel_request(req);
2613 	linger_put(lreq);
2614 }
2615 
2616 struct linger_work {
2617 	struct work_struct work;
2618 	struct ceph_osd_linger_request *lreq;
2619 	struct list_head pending_item;
2620 	unsigned long queued_stamp;
2621 
2622 	union {
2623 		struct {
2624 			u64 notify_id;
2625 			u64 notifier_id;
2626 			void *payload; /* points into @msg front */
2627 			size_t payload_len;
2628 
2629 			struct ceph_msg *msg; /* for ceph_msg_put() */
2630 		} notify;
2631 		struct {
2632 			int err;
2633 		} error;
2634 	};
2635 };
2636 
2637 static struct linger_work *lwork_alloc(struct ceph_osd_linger_request *lreq,
2638 				       work_func_t workfn)
2639 {
2640 	struct linger_work *lwork;
2641 
2642 	lwork = kzalloc(sizeof(*lwork), GFP_NOIO);
2643 	if (!lwork)
2644 		return NULL;
2645 
2646 	INIT_WORK(&lwork->work, workfn);
2647 	INIT_LIST_HEAD(&lwork->pending_item);
2648 	lwork->lreq = linger_get(lreq);
2649 
2650 	return lwork;
2651 }
2652 
2653 static void lwork_free(struct linger_work *lwork)
2654 {
2655 	struct ceph_osd_linger_request *lreq = lwork->lreq;
2656 
2657 	mutex_lock(&lreq->lock);
2658 	list_del(&lwork->pending_item);
2659 	mutex_unlock(&lreq->lock);
2660 
2661 	linger_put(lreq);
2662 	kfree(lwork);
2663 }
2664 
2665 static void lwork_queue(struct linger_work *lwork)
2666 {
2667 	struct ceph_osd_linger_request *lreq = lwork->lreq;
2668 	struct ceph_osd_client *osdc = lreq->osdc;
2669 
2670 	verify_lreq_locked(lreq);
2671 	WARN_ON(!list_empty(&lwork->pending_item));
2672 
2673 	lwork->queued_stamp = jiffies;
2674 	list_add_tail(&lwork->pending_item, &lreq->pending_lworks);
2675 	queue_work(osdc->notify_wq, &lwork->work);
2676 }
2677 
2678 static void do_watch_notify(struct work_struct *w)
2679 {
2680 	struct linger_work *lwork = container_of(w, struct linger_work, work);
2681 	struct ceph_osd_linger_request *lreq = lwork->lreq;
2682 
2683 	if (!linger_registered(lreq)) {
2684 		dout("%s lreq %p not registered\n", __func__, lreq);
2685 		goto out;
2686 	}
2687 
2688 	WARN_ON(!lreq->is_watch);
2689 	dout("%s lreq %p notify_id %llu notifier_id %llu payload_len %zu\n",
2690 	     __func__, lreq, lwork->notify.notify_id, lwork->notify.notifier_id,
2691 	     lwork->notify.payload_len);
2692 	lreq->wcb(lreq->data, lwork->notify.notify_id, lreq->linger_id,
2693 		  lwork->notify.notifier_id, lwork->notify.payload,
2694 		  lwork->notify.payload_len);
2695 
2696 out:
2697 	ceph_msg_put(lwork->notify.msg);
2698 	lwork_free(lwork);
2699 }
2700 
2701 static void do_watch_error(struct work_struct *w)
2702 {
2703 	struct linger_work *lwork = container_of(w, struct linger_work, work);
2704 	struct ceph_osd_linger_request *lreq = lwork->lreq;
2705 
2706 	if (!linger_registered(lreq)) {
2707 		dout("%s lreq %p not registered\n", __func__, lreq);
2708 		goto out;
2709 	}
2710 
2711 	dout("%s lreq %p err %d\n", __func__, lreq, lwork->error.err);
2712 	lreq->errcb(lreq->data, lreq->linger_id, lwork->error.err);
2713 
2714 out:
2715 	lwork_free(lwork);
2716 }
2717 
2718 static void queue_watch_error(struct ceph_osd_linger_request *lreq)
2719 {
2720 	struct linger_work *lwork;
2721 
2722 	lwork = lwork_alloc(lreq, do_watch_error);
2723 	if (!lwork) {
2724 		pr_err("failed to allocate error-lwork\n");
2725 		return;
2726 	}
2727 
2728 	lwork->error.err = lreq->last_error;
2729 	lwork_queue(lwork);
2730 }
2731 
2732 static void linger_reg_commit_complete(struct ceph_osd_linger_request *lreq,
2733 				       int result)
2734 {
2735 	if (!completion_done(&lreq->reg_commit_wait)) {
2736 		lreq->reg_commit_error = (result <= 0 ? result : 0);
2737 		complete_all(&lreq->reg_commit_wait);
2738 	}
2739 }
2740 
2741 static void linger_commit_cb(struct ceph_osd_request *req)
2742 {
2743 	struct ceph_osd_linger_request *lreq = req->r_priv;
2744 
2745 	mutex_lock(&lreq->lock);
2746 	dout("%s lreq %p linger_id %llu result %d\n", __func__, lreq,
2747 	     lreq->linger_id, req->r_result);
2748 	linger_reg_commit_complete(lreq, req->r_result);
2749 	lreq->committed = true;
2750 
2751 	if (!lreq->is_watch) {
2752 		struct ceph_osd_data *osd_data =
2753 		    osd_req_op_data(req, 0, notify, response_data);
2754 		void *p = page_address(osd_data->pages[0]);
2755 
2756 		WARN_ON(req->r_ops[0].op != CEPH_OSD_OP_NOTIFY ||
2757 			osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
2758 
2759 		/* make note of the notify_id */
2760 		if (req->r_ops[0].outdata_len >= sizeof(u64)) {
2761 			lreq->notify_id = ceph_decode_64(&p);
2762 			dout("lreq %p notify_id %llu\n", lreq,
2763 			     lreq->notify_id);
2764 		} else {
2765 			dout("lreq %p no notify_id\n", lreq);
2766 		}
2767 	}
2768 
2769 	mutex_unlock(&lreq->lock);
2770 	linger_put(lreq);
2771 }
2772 
2773 static int normalize_watch_error(int err)
2774 {
2775 	/*
2776 	 * Translate ENOENT -> ENOTCONN so that a delete->disconnection
2777 	 * notification and a failure to reconnect because we raced with
2778 	 * the delete appear the same to the user.
2779 	 */
2780 	if (err == -ENOENT)
2781 		err = -ENOTCONN;
2782 
2783 	return err;
2784 }
2785 
2786 static void linger_reconnect_cb(struct ceph_osd_request *req)
2787 {
2788 	struct ceph_osd_linger_request *lreq = req->r_priv;
2789 
2790 	mutex_lock(&lreq->lock);
2791 	dout("%s lreq %p linger_id %llu result %d last_error %d\n", __func__,
2792 	     lreq, lreq->linger_id, req->r_result, lreq->last_error);
2793 	if (req->r_result < 0) {
2794 		if (!lreq->last_error) {
2795 			lreq->last_error = normalize_watch_error(req->r_result);
2796 			queue_watch_error(lreq);
2797 		}
2798 	}
2799 
2800 	mutex_unlock(&lreq->lock);
2801 	linger_put(lreq);
2802 }
2803 
2804 static void send_linger(struct ceph_osd_linger_request *lreq)
2805 {
2806 	struct ceph_osd_request *req = lreq->reg_req;
2807 	struct ceph_osd_req_op *op = &req->r_ops[0];
2808 
2809 	verify_osdc_wrlocked(req->r_osdc);
2810 	dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id);
2811 
2812 	if (req->r_osd)
2813 		cancel_linger_request(req);
2814 
2815 	request_reinit(req);
2816 	ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid);
2817 	ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc);
2818 	req->r_flags = lreq->t.flags;
2819 	req->r_mtime = lreq->mtime;
2820 
2821 	mutex_lock(&lreq->lock);
2822 	if (lreq->is_watch && lreq->committed) {
2823 		WARN_ON(op->op != CEPH_OSD_OP_WATCH ||
2824 			op->watch.cookie != lreq->linger_id);
2825 		op->watch.op = CEPH_OSD_WATCH_OP_RECONNECT;
2826 		op->watch.gen = ++lreq->register_gen;
2827 		dout("lreq %p reconnect register_gen %u\n", lreq,
2828 		     op->watch.gen);
2829 		req->r_callback = linger_reconnect_cb;
2830 	} else {
2831 		if (!lreq->is_watch)
2832 			lreq->notify_id = 0;
2833 		else
2834 			WARN_ON(op->watch.op != CEPH_OSD_WATCH_OP_WATCH);
2835 		dout("lreq %p register\n", lreq);
2836 		req->r_callback = linger_commit_cb;
2837 	}
2838 	mutex_unlock(&lreq->lock);
2839 
2840 	req->r_priv = linger_get(lreq);
2841 	req->r_linger = true;
2842 
2843 	submit_request(req, true);
2844 }
2845 
2846 static void linger_ping_cb(struct ceph_osd_request *req)
2847 {
2848 	struct ceph_osd_linger_request *lreq = req->r_priv;
2849 
2850 	mutex_lock(&lreq->lock);
2851 	dout("%s lreq %p linger_id %llu result %d ping_sent %lu last_error %d\n",
2852 	     __func__, lreq, lreq->linger_id, req->r_result, lreq->ping_sent,
2853 	     lreq->last_error);
2854 	if (lreq->register_gen == req->r_ops[0].watch.gen) {
2855 		if (!req->r_result) {
2856 			lreq->watch_valid_thru = lreq->ping_sent;
2857 		} else if (!lreq->last_error) {
2858 			lreq->last_error = normalize_watch_error(req->r_result);
2859 			queue_watch_error(lreq);
2860 		}
2861 	} else {
2862 		dout("lreq %p register_gen %u ignoring old pong %u\n", lreq,
2863 		     lreq->register_gen, req->r_ops[0].watch.gen);
2864 	}
2865 
2866 	mutex_unlock(&lreq->lock);
2867 	linger_put(lreq);
2868 }
2869 
2870 static void send_linger_ping(struct ceph_osd_linger_request *lreq)
2871 {
2872 	struct ceph_osd_client *osdc = lreq->osdc;
2873 	struct ceph_osd_request *req = lreq->ping_req;
2874 	struct ceph_osd_req_op *op = &req->r_ops[0];
2875 
2876 	if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD)) {
2877 		dout("%s PAUSERD\n", __func__);
2878 		return;
2879 	}
2880 
2881 	lreq->ping_sent = jiffies;
2882 	dout("%s lreq %p linger_id %llu ping_sent %lu register_gen %u\n",
2883 	     __func__, lreq, lreq->linger_id, lreq->ping_sent,
2884 	     lreq->register_gen);
2885 
2886 	if (req->r_osd)
2887 		cancel_linger_request(req);
2888 
2889 	request_reinit(req);
2890 	target_copy(&req->r_t, &lreq->t);
2891 
2892 	WARN_ON(op->op != CEPH_OSD_OP_WATCH ||
2893 		op->watch.cookie != lreq->linger_id ||
2894 		op->watch.op != CEPH_OSD_WATCH_OP_PING);
2895 	op->watch.gen = lreq->register_gen;
2896 	req->r_callback = linger_ping_cb;
2897 	req->r_priv = linger_get(lreq);
2898 	req->r_linger = true;
2899 
2900 	ceph_osdc_get_request(req);
2901 	account_request(req);
2902 	req->r_tid = atomic64_inc_return(&osdc->last_tid);
2903 	link_request(lreq->osd, req);
2904 	send_request(req);
2905 }
2906 
2907 static void linger_submit(struct ceph_osd_linger_request *lreq)
2908 {
2909 	struct ceph_osd_client *osdc = lreq->osdc;
2910 	struct ceph_osd *osd;
2911 
2912 	calc_target(osdc, &lreq->t, NULL, false);
2913 	osd = lookup_create_osd(osdc, lreq->t.osd, true);
2914 	link_linger(osd, lreq);
2915 
2916 	send_linger(lreq);
2917 }
2918 
2919 static void cancel_linger_map_check(struct ceph_osd_linger_request *lreq)
2920 {
2921 	struct ceph_osd_client *osdc = lreq->osdc;
2922 	struct ceph_osd_linger_request *lookup_lreq;
2923 
2924 	verify_osdc_wrlocked(osdc);
2925 
2926 	lookup_lreq = lookup_linger_mc(&osdc->linger_map_checks,
2927 				       lreq->linger_id);
2928 	if (!lookup_lreq)
2929 		return;
2930 
2931 	WARN_ON(lookup_lreq != lreq);
2932 	erase_linger_mc(&osdc->linger_map_checks, lreq);
2933 	linger_put(lreq);
2934 }
2935 
2936 /*
2937  * @lreq has to be both registered and linked.
2938  */
2939 static void __linger_cancel(struct ceph_osd_linger_request *lreq)
2940 {
2941 	if (lreq->is_watch && lreq->ping_req->r_osd)
2942 		cancel_linger_request(lreq->ping_req);
2943 	if (lreq->reg_req->r_osd)
2944 		cancel_linger_request(lreq->reg_req);
2945 	cancel_linger_map_check(lreq);
2946 	unlink_linger(lreq->osd, lreq);
2947 	linger_unregister(lreq);
2948 }
2949 
2950 static void linger_cancel(struct ceph_osd_linger_request *lreq)
2951 {
2952 	struct ceph_osd_client *osdc = lreq->osdc;
2953 
2954 	down_write(&osdc->lock);
2955 	if (__linger_registered(lreq))
2956 		__linger_cancel(lreq);
2957 	up_write(&osdc->lock);
2958 }
2959 
2960 static void send_linger_map_check(struct ceph_osd_linger_request *lreq);
2961 
2962 static void check_linger_pool_dne(struct ceph_osd_linger_request *lreq)
2963 {
2964 	struct ceph_osd_client *osdc = lreq->osdc;
2965 	struct ceph_osdmap *map = osdc->osdmap;
2966 
2967 	verify_osdc_wrlocked(osdc);
2968 	WARN_ON(!map->epoch);
2969 
2970 	if (lreq->register_gen) {
2971 		lreq->map_dne_bound = map->epoch;
2972 		dout("%s lreq %p linger_id %llu pool disappeared\n", __func__,
2973 		     lreq, lreq->linger_id);
2974 	} else {
2975 		dout("%s lreq %p linger_id %llu map_dne_bound %u have %u\n",
2976 		     __func__, lreq, lreq->linger_id, lreq->map_dne_bound,
2977 		     map->epoch);
2978 	}
2979 
2980 	if (lreq->map_dne_bound) {
2981 		if (map->epoch >= lreq->map_dne_bound) {
2982 			/* we had a new enough map */
2983 			pr_info("linger_id %llu pool does not exist\n",
2984 				lreq->linger_id);
2985 			linger_reg_commit_complete(lreq, -ENOENT);
2986 			__linger_cancel(lreq);
2987 		}
2988 	} else {
2989 		send_linger_map_check(lreq);
2990 	}
2991 }
2992 
2993 static void linger_map_check_cb(struct ceph_mon_generic_request *greq)
2994 {
2995 	struct ceph_osd_client *osdc = &greq->monc->client->osdc;
2996 	struct ceph_osd_linger_request *lreq;
2997 	u64 linger_id = greq->private_data;
2998 
2999 	WARN_ON(greq->result || !greq->u.newest);
3000 
3001 	down_write(&osdc->lock);
3002 	lreq = lookup_linger_mc(&osdc->linger_map_checks, linger_id);
3003 	if (!lreq) {
3004 		dout("%s linger_id %llu dne\n", __func__, linger_id);
3005 		goto out_unlock;
3006 	}
3007 
3008 	dout("%s lreq %p linger_id %llu map_dne_bound %u newest %llu\n",
3009 	     __func__, lreq, lreq->linger_id, lreq->map_dne_bound,
3010 	     greq->u.newest);
3011 	if (!lreq->map_dne_bound)
3012 		lreq->map_dne_bound = greq->u.newest;
3013 	erase_linger_mc(&osdc->linger_map_checks, lreq);
3014 	check_linger_pool_dne(lreq);
3015 
3016 	linger_put(lreq);
3017 out_unlock:
3018 	up_write(&osdc->lock);
3019 }
3020 
3021 static void send_linger_map_check(struct ceph_osd_linger_request *lreq)
3022 {
3023 	struct ceph_osd_client *osdc = lreq->osdc;
3024 	struct ceph_osd_linger_request *lookup_lreq;
3025 	int ret;
3026 
3027 	verify_osdc_wrlocked(osdc);
3028 
3029 	lookup_lreq = lookup_linger_mc(&osdc->linger_map_checks,
3030 				       lreq->linger_id);
3031 	if (lookup_lreq) {
3032 		WARN_ON(lookup_lreq != lreq);
3033 		return;
3034 	}
3035 
3036 	linger_get(lreq);
3037 	insert_linger_mc(&osdc->linger_map_checks, lreq);
3038 	ret = ceph_monc_get_version_async(&osdc->client->monc, "osdmap",
3039 					  linger_map_check_cb, lreq->linger_id);
3040 	WARN_ON(ret);
3041 }
3042 
3043 static int linger_reg_commit_wait(struct ceph_osd_linger_request *lreq)
3044 {
3045 	int ret;
3046 
3047 	dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id);
3048 	ret = wait_for_completion_interruptible(&lreq->reg_commit_wait);
3049 	return ret ?: lreq->reg_commit_error;
3050 }
3051 
3052 static int linger_notify_finish_wait(struct ceph_osd_linger_request *lreq)
3053 {
3054 	int ret;
3055 
3056 	dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id);
3057 	ret = wait_for_completion_interruptible(&lreq->notify_finish_wait);
3058 	return ret ?: lreq->notify_finish_error;
3059 }
3060 
3061 /*
3062  * Timeout callback, called every N seconds.  When 1 or more OSD
3063  * requests has been active for more than N seconds, we send a keepalive
3064  * (tag + timestamp) to its OSD to ensure any communications channel
3065  * reset is detected.
3066  */
3067 static void handle_timeout(struct work_struct *work)
3068 {
3069 	struct ceph_osd_client *osdc =
3070 		container_of(work, struct ceph_osd_client, timeout_work.work);
3071 	struct ceph_options *opts = osdc->client->options;
3072 	unsigned long cutoff = jiffies - opts->osd_keepalive_timeout;
3073 	unsigned long expiry_cutoff = jiffies - opts->osd_request_timeout;
3074 	LIST_HEAD(slow_osds);
3075 	struct rb_node *n, *p;
3076 
3077 	dout("%s osdc %p\n", __func__, osdc);
3078 	down_write(&osdc->lock);
3079 
3080 	/*
3081 	 * ping osds that are a bit slow.  this ensures that if there
3082 	 * is a break in the TCP connection we will notice, and reopen
3083 	 * a connection with that osd (from the fault callback).
3084 	 */
3085 	for (n = rb_first(&osdc->osds); n; n = rb_next(n)) {
3086 		struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node);
3087 		bool found = false;
3088 
3089 		for (p = rb_first(&osd->o_requests); p; ) {
3090 			struct ceph_osd_request *req =
3091 			    rb_entry(p, struct ceph_osd_request, r_node);
3092 
3093 			p = rb_next(p); /* abort_request() */
3094 
3095 			if (time_before(req->r_stamp, cutoff)) {
3096 				dout(" req %p tid %llu on osd%d is laggy\n",
3097 				     req, req->r_tid, osd->o_osd);
3098 				found = true;
3099 			}
3100 			if (opts->osd_request_timeout &&
3101 			    time_before(req->r_start_stamp, expiry_cutoff)) {
3102 				pr_err_ratelimited("tid %llu on osd%d timeout\n",
3103 				       req->r_tid, osd->o_osd);
3104 				abort_request(req, -ETIMEDOUT);
3105 			}
3106 		}
3107 		for (p = rb_first(&osd->o_linger_requests); p; p = rb_next(p)) {
3108 			struct ceph_osd_linger_request *lreq =
3109 			    rb_entry(p, struct ceph_osd_linger_request, node);
3110 
3111 			dout(" lreq %p linger_id %llu is served by osd%d\n",
3112 			     lreq, lreq->linger_id, osd->o_osd);
3113 			found = true;
3114 
3115 			mutex_lock(&lreq->lock);
3116 			if (lreq->is_watch && lreq->committed && !lreq->last_error)
3117 				send_linger_ping(lreq);
3118 			mutex_unlock(&lreq->lock);
3119 		}
3120 
3121 		if (found)
3122 			list_move_tail(&osd->o_keepalive_item, &slow_osds);
3123 	}
3124 
3125 	if (opts->osd_request_timeout) {
3126 		for (p = rb_first(&osdc->homeless_osd.o_requests); p; ) {
3127 			struct ceph_osd_request *req =
3128 			    rb_entry(p, struct ceph_osd_request, r_node);
3129 
3130 			p = rb_next(p); /* abort_request() */
3131 
3132 			if (time_before(req->r_start_stamp, expiry_cutoff)) {
3133 				pr_err_ratelimited("tid %llu on osd%d timeout\n",
3134 				       req->r_tid, osdc->homeless_osd.o_osd);
3135 				abort_request(req, -ETIMEDOUT);
3136 			}
3137 		}
3138 	}
3139 
3140 	if (atomic_read(&osdc->num_homeless) || !list_empty(&slow_osds))
3141 		maybe_request_map(osdc);
3142 
3143 	while (!list_empty(&slow_osds)) {
3144 		struct ceph_osd *osd = list_first_entry(&slow_osds,
3145 							struct ceph_osd,
3146 							o_keepalive_item);
3147 		list_del_init(&osd->o_keepalive_item);
3148 		ceph_con_keepalive(&osd->o_con);
3149 	}
3150 
3151 	up_write(&osdc->lock);
3152 	schedule_delayed_work(&osdc->timeout_work,
3153 			      osdc->client->options->osd_keepalive_timeout);
3154 }
3155 
3156 static void handle_osds_timeout(struct work_struct *work)
3157 {
3158 	struct ceph_osd_client *osdc =
3159 		container_of(work, struct ceph_osd_client,
3160 			     osds_timeout_work.work);
3161 	unsigned long delay = osdc->client->options->osd_idle_ttl / 4;
3162 	struct ceph_osd *osd, *nosd;
3163 
3164 	dout("%s osdc %p\n", __func__, osdc);
3165 	down_write(&osdc->lock);
3166 	list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
3167 		if (time_before(jiffies, osd->lru_ttl))
3168 			break;
3169 
3170 		WARN_ON(!RB_EMPTY_ROOT(&osd->o_requests));
3171 		WARN_ON(!RB_EMPTY_ROOT(&osd->o_linger_requests));
3172 		close_osd(osd);
3173 	}
3174 
3175 	up_write(&osdc->lock);
3176 	schedule_delayed_work(&osdc->osds_timeout_work,
3177 			      round_jiffies_relative(delay));
3178 }
3179 
3180 static int ceph_oloc_decode(void **p, void *end,
3181 			    struct ceph_object_locator *oloc)
3182 {
3183 	u8 struct_v, struct_cv;
3184 	u32 len;
3185 	void *struct_end;
3186 	int ret = 0;
3187 
3188 	ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
3189 	struct_v = ceph_decode_8(p);
3190 	struct_cv = ceph_decode_8(p);
3191 	if (struct_v < 3) {
3192 		pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
3193 			struct_v, struct_cv);
3194 		goto e_inval;
3195 	}
3196 	if (struct_cv > 6) {
3197 		pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
3198 			struct_v, struct_cv);
3199 		goto e_inval;
3200 	}
3201 	len = ceph_decode_32(p);
3202 	ceph_decode_need(p, end, len, e_inval);
3203 	struct_end = *p + len;
3204 
3205 	oloc->pool = ceph_decode_64(p);
3206 	*p += 4; /* skip preferred */
3207 
3208 	len = ceph_decode_32(p);
3209 	if (len > 0) {
3210 		pr_warn("ceph_object_locator::key is set\n");
3211 		goto e_inval;
3212 	}
3213 
3214 	if (struct_v >= 5) {
3215 		bool changed = false;
3216 
3217 		len = ceph_decode_32(p);
3218 		if (len > 0) {
3219 			ceph_decode_need(p, end, len, e_inval);
3220 			if (!oloc->pool_ns ||
3221 			    ceph_compare_string(oloc->pool_ns, *p, len))
3222 				changed = true;
3223 			*p += len;
3224 		} else {
3225 			if (oloc->pool_ns)
3226 				changed = true;
3227 		}
3228 		if (changed) {
3229 			/* redirect changes namespace */
3230 			pr_warn("ceph_object_locator::nspace is changed\n");
3231 			goto e_inval;
3232 		}
3233 	}
3234 
3235 	if (struct_v >= 6) {
3236 		s64 hash = ceph_decode_64(p);
3237 		if (hash != -1) {
3238 			pr_warn("ceph_object_locator::hash is set\n");
3239 			goto e_inval;
3240 		}
3241 	}
3242 
3243 	/* skip the rest */
3244 	*p = struct_end;
3245 out:
3246 	return ret;
3247 
3248 e_inval:
3249 	ret = -EINVAL;
3250 	goto out;
3251 }
3252 
3253 static int ceph_redirect_decode(void **p, void *end,
3254 				struct ceph_request_redirect *redir)
3255 {
3256 	u8 struct_v, struct_cv;
3257 	u32 len;
3258 	void *struct_end;
3259 	int ret;
3260 
3261 	ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
3262 	struct_v = ceph_decode_8(p);
3263 	struct_cv = ceph_decode_8(p);
3264 	if (struct_cv > 1) {
3265 		pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
3266 			struct_v, struct_cv);
3267 		goto e_inval;
3268 	}
3269 	len = ceph_decode_32(p);
3270 	ceph_decode_need(p, end, len, e_inval);
3271 	struct_end = *p + len;
3272 
3273 	ret = ceph_oloc_decode(p, end, &redir->oloc);
3274 	if (ret)
3275 		goto out;
3276 
3277 	len = ceph_decode_32(p);
3278 	if (len > 0) {
3279 		pr_warn("ceph_request_redirect::object_name is set\n");
3280 		goto e_inval;
3281 	}
3282 
3283 	len = ceph_decode_32(p);
3284 	*p += len; /* skip osd_instructions */
3285 
3286 	/* skip the rest */
3287 	*p = struct_end;
3288 out:
3289 	return ret;
3290 
3291 e_inval:
3292 	ret = -EINVAL;
3293 	goto out;
3294 }
3295 
3296 struct MOSDOpReply {
3297 	struct ceph_pg pgid;
3298 	u64 flags;
3299 	int result;
3300 	u32 epoch;
3301 	int num_ops;
3302 	u32 outdata_len[CEPH_OSD_MAX_OPS];
3303 	s32 rval[CEPH_OSD_MAX_OPS];
3304 	int retry_attempt;
3305 	struct ceph_eversion replay_version;
3306 	u64 user_version;
3307 	struct ceph_request_redirect redirect;
3308 };
3309 
3310 static int decode_MOSDOpReply(const struct ceph_msg *msg, struct MOSDOpReply *m)
3311 {
3312 	void *p = msg->front.iov_base;
3313 	void *const end = p + msg->front.iov_len;
3314 	u16 version = le16_to_cpu(msg->hdr.version);
3315 	struct ceph_eversion bad_replay_version;
3316 	u8 decode_redir;
3317 	u32 len;
3318 	int ret;
3319 	int i;
3320 
3321 	ceph_decode_32_safe(&p, end, len, e_inval);
3322 	ceph_decode_need(&p, end, len, e_inval);
3323 	p += len; /* skip oid */
3324 
3325 	ret = ceph_decode_pgid(&p, end, &m->pgid);
3326 	if (ret)
3327 		return ret;
3328 
3329 	ceph_decode_64_safe(&p, end, m->flags, e_inval);
3330 	ceph_decode_32_safe(&p, end, m->result, e_inval);
3331 	ceph_decode_need(&p, end, sizeof(bad_replay_version), e_inval);
3332 	memcpy(&bad_replay_version, p, sizeof(bad_replay_version));
3333 	p += sizeof(bad_replay_version);
3334 	ceph_decode_32_safe(&p, end, m->epoch, e_inval);
3335 
3336 	ceph_decode_32_safe(&p, end, m->num_ops, e_inval);
3337 	if (m->num_ops > ARRAY_SIZE(m->outdata_len))
3338 		goto e_inval;
3339 
3340 	ceph_decode_need(&p, end, m->num_ops * sizeof(struct ceph_osd_op),
3341 			 e_inval);
3342 	for (i = 0; i < m->num_ops; i++) {
3343 		struct ceph_osd_op *op = p;
3344 
3345 		m->outdata_len[i] = le32_to_cpu(op->payload_len);
3346 		p += sizeof(*op);
3347 	}
3348 
3349 	ceph_decode_32_safe(&p, end, m->retry_attempt, e_inval);
3350 	for (i = 0; i < m->num_ops; i++)
3351 		ceph_decode_32_safe(&p, end, m->rval[i], e_inval);
3352 
3353 	if (version >= 5) {
3354 		ceph_decode_need(&p, end, sizeof(m->replay_version), e_inval);
3355 		memcpy(&m->replay_version, p, sizeof(m->replay_version));
3356 		p += sizeof(m->replay_version);
3357 		ceph_decode_64_safe(&p, end, m->user_version, e_inval);
3358 	} else {
3359 		m->replay_version = bad_replay_version; /* struct */
3360 		m->user_version = le64_to_cpu(m->replay_version.version);
3361 	}
3362 
3363 	if (version >= 6) {
3364 		if (version >= 7)
3365 			ceph_decode_8_safe(&p, end, decode_redir, e_inval);
3366 		else
3367 			decode_redir = 1;
3368 	} else {
3369 		decode_redir = 0;
3370 	}
3371 
3372 	if (decode_redir) {
3373 		ret = ceph_redirect_decode(&p, end, &m->redirect);
3374 		if (ret)
3375 			return ret;
3376 	} else {
3377 		ceph_oloc_init(&m->redirect.oloc);
3378 	}
3379 
3380 	return 0;
3381 
3382 e_inval:
3383 	return -EINVAL;
3384 }
3385 
3386 /*
3387  * Handle MOSDOpReply.  Set ->r_result and call the callback if it is
3388  * specified.
3389  */
3390 static void handle_reply(struct ceph_osd *osd, struct ceph_msg *msg)
3391 {
3392 	struct ceph_osd_client *osdc = osd->o_osdc;
3393 	struct ceph_osd_request *req;
3394 	struct MOSDOpReply m;
3395 	u64 tid = le64_to_cpu(msg->hdr.tid);
3396 	u32 data_len = 0;
3397 	int ret;
3398 	int i;
3399 
3400 	dout("%s msg %p tid %llu\n", __func__, msg, tid);
3401 
3402 	down_read(&osdc->lock);
3403 	if (!osd_registered(osd)) {
3404 		dout("%s osd%d unknown\n", __func__, osd->o_osd);
3405 		goto out_unlock_osdc;
3406 	}
3407 	WARN_ON(osd->o_osd != le64_to_cpu(msg->hdr.src.num));
3408 
3409 	mutex_lock(&osd->lock);
3410 	req = lookup_request(&osd->o_requests, tid);
3411 	if (!req) {
3412 		dout("%s osd%d tid %llu unknown\n", __func__, osd->o_osd, tid);
3413 		goto out_unlock_session;
3414 	}
3415 
3416 	m.redirect.oloc.pool_ns = req->r_t.target_oloc.pool_ns;
3417 	ret = decode_MOSDOpReply(msg, &m);
3418 	m.redirect.oloc.pool_ns = NULL;
3419 	if (ret) {
3420 		pr_err("failed to decode MOSDOpReply for tid %llu: %d\n",
3421 		       req->r_tid, ret);
3422 		ceph_msg_dump(msg);
3423 		goto fail_request;
3424 	}
3425 	dout("%s req %p tid %llu flags 0x%llx pgid %llu.%x epoch %u attempt %d v %u'%llu uv %llu\n",
3426 	     __func__, req, req->r_tid, m.flags, m.pgid.pool, m.pgid.seed,
3427 	     m.epoch, m.retry_attempt, le32_to_cpu(m.replay_version.epoch),
3428 	     le64_to_cpu(m.replay_version.version), m.user_version);
3429 
3430 	if (m.retry_attempt >= 0) {
3431 		if (m.retry_attempt != req->r_attempts - 1) {
3432 			dout("req %p tid %llu retry_attempt %d != %d, ignoring\n",
3433 			     req, req->r_tid, m.retry_attempt,
3434 			     req->r_attempts - 1);
3435 			goto out_unlock_session;
3436 		}
3437 	} else {
3438 		WARN_ON(1); /* MOSDOpReply v4 is assumed */
3439 	}
3440 
3441 	if (!ceph_oloc_empty(&m.redirect.oloc)) {
3442 		dout("req %p tid %llu redirect pool %lld\n", req, req->r_tid,
3443 		     m.redirect.oloc.pool);
3444 		unlink_request(osd, req);
3445 		mutex_unlock(&osd->lock);
3446 
3447 		/*
3448 		 * Not ceph_oloc_copy() - changing pool_ns is not
3449 		 * supported.
3450 		 */
3451 		req->r_t.target_oloc.pool = m.redirect.oloc.pool;
3452 		req->r_flags |= CEPH_OSD_FLAG_REDIRECTED;
3453 		req->r_tid = 0;
3454 		__submit_request(req, false);
3455 		goto out_unlock_osdc;
3456 	}
3457 
3458 	if (m.num_ops != req->r_num_ops) {
3459 		pr_err("num_ops %d != %d for tid %llu\n", m.num_ops,
3460 		       req->r_num_ops, req->r_tid);
3461 		goto fail_request;
3462 	}
3463 	for (i = 0; i < req->r_num_ops; i++) {
3464 		dout(" req %p tid %llu op %d rval %d len %u\n", req,
3465 		     req->r_tid, i, m.rval[i], m.outdata_len[i]);
3466 		req->r_ops[i].rval = m.rval[i];
3467 		req->r_ops[i].outdata_len = m.outdata_len[i];
3468 		data_len += m.outdata_len[i];
3469 	}
3470 	if (data_len != le32_to_cpu(msg->hdr.data_len)) {
3471 		pr_err("sum of lens %u != %u for tid %llu\n", data_len,
3472 		       le32_to_cpu(msg->hdr.data_len), req->r_tid);
3473 		goto fail_request;
3474 	}
3475 	dout("%s req %p tid %llu result %d data_len %u\n", __func__,
3476 	     req, req->r_tid, m.result, data_len);
3477 
3478 	/*
3479 	 * Since we only ever request ONDISK, we should only ever get
3480 	 * one (type of) reply back.
3481 	 */
3482 	WARN_ON(!(m.flags & CEPH_OSD_FLAG_ONDISK));
3483 	req->r_result = m.result ?: data_len;
3484 	finish_request(req);
3485 	mutex_unlock(&osd->lock);
3486 	up_read(&osdc->lock);
3487 
3488 	__complete_request(req);
3489 	complete_all(&req->r_completion);
3490 	ceph_osdc_put_request(req);
3491 	return;
3492 
3493 fail_request:
3494 	complete_request(req, -EIO);
3495 out_unlock_session:
3496 	mutex_unlock(&osd->lock);
3497 out_unlock_osdc:
3498 	up_read(&osdc->lock);
3499 }
3500 
3501 static void set_pool_was_full(struct ceph_osd_client *osdc)
3502 {
3503 	struct rb_node *n;
3504 
3505 	for (n = rb_first(&osdc->osdmap->pg_pools); n; n = rb_next(n)) {
3506 		struct ceph_pg_pool_info *pi =
3507 		    rb_entry(n, struct ceph_pg_pool_info, node);
3508 
3509 		pi->was_full = __pool_full(pi);
3510 	}
3511 }
3512 
3513 static bool pool_cleared_full(struct ceph_osd_client *osdc, s64 pool_id)
3514 {
3515 	struct ceph_pg_pool_info *pi;
3516 
3517 	pi = ceph_pg_pool_by_id(osdc->osdmap, pool_id);
3518 	if (!pi)
3519 		return false;
3520 
3521 	return pi->was_full && !__pool_full(pi);
3522 }
3523 
3524 static enum calc_target_result
3525 recalc_linger_target(struct ceph_osd_linger_request *lreq)
3526 {
3527 	struct ceph_osd_client *osdc = lreq->osdc;
3528 	enum calc_target_result ct_res;
3529 
3530 	ct_res = calc_target(osdc, &lreq->t, NULL, true);
3531 	if (ct_res == CALC_TARGET_NEED_RESEND) {
3532 		struct ceph_osd *osd;
3533 
3534 		osd = lookup_create_osd(osdc, lreq->t.osd, true);
3535 		if (osd != lreq->osd) {
3536 			unlink_linger(lreq->osd, lreq);
3537 			link_linger(osd, lreq);
3538 		}
3539 	}
3540 
3541 	return ct_res;
3542 }
3543 
3544 /*
3545  * Requeue requests whose mapping to an OSD has changed.
3546  */
3547 static void scan_requests(struct ceph_osd *osd,
3548 			  bool force_resend,
3549 			  bool cleared_full,
3550 			  bool check_pool_cleared_full,
3551 			  struct rb_root *need_resend,
3552 			  struct list_head *need_resend_linger)
3553 {
3554 	struct ceph_osd_client *osdc = osd->o_osdc;
3555 	struct rb_node *n;
3556 	bool force_resend_writes;
3557 
3558 	for (n = rb_first(&osd->o_linger_requests); n; ) {
3559 		struct ceph_osd_linger_request *lreq =
3560 		    rb_entry(n, struct ceph_osd_linger_request, node);
3561 		enum calc_target_result ct_res;
3562 
3563 		n = rb_next(n); /* recalc_linger_target() */
3564 
3565 		dout("%s lreq %p linger_id %llu\n", __func__, lreq,
3566 		     lreq->linger_id);
3567 		ct_res = recalc_linger_target(lreq);
3568 		switch (ct_res) {
3569 		case CALC_TARGET_NO_ACTION:
3570 			force_resend_writes = cleared_full ||
3571 			    (check_pool_cleared_full &&
3572 			     pool_cleared_full(osdc, lreq->t.base_oloc.pool));
3573 			if (!force_resend && !force_resend_writes)
3574 				break;
3575 
3576 			/* fall through */
3577 		case CALC_TARGET_NEED_RESEND:
3578 			cancel_linger_map_check(lreq);
3579 			/*
3580 			 * scan_requests() for the previous epoch(s)
3581 			 * may have already added it to the list, since
3582 			 * it's not unlinked here.
3583 			 */
3584 			if (list_empty(&lreq->scan_item))
3585 				list_add_tail(&lreq->scan_item, need_resend_linger);
3586 			break;
3587 		case CALC_TARGET_POOL_DNE:
3588 			list_del_init(&lreq->scan_item);
3589 			check_linger_pool_dne(lreq);
3590 			break;
3591 		}
3592 	}
3593 
3594 	for (n = rb_first(&osd->o_requests); n; ) {
3595 		struct ceph_osd_request *req =
3596 		    rb_entry(n, struct ceph_osd_request, r_node);
3597 		enum calc_target_result ct_res;
3598 
3599 		n = rb_next(n); /* unlink_request(), check_pool_dne() */
3600 
3601 		dout("%s req %p tid %llu\n", __func__, req, req->r_tid);
3602 		ct_res = calc_target(osdc, &req->r_t, &req->r_osd->o_con,
3603 				     false);
3604 		switch (ct_res) {
3605 		case CALC_TARGET_NO_ACTION:
3606 			force_resend_writes = cleared_full ||
3607 			    (check_pool_cleared_full &&
3608 			     pool_cleared_full(osdc, req->r_t.base_oloc.pool));
3609 			if (!force_resend &&
3610 			    (!(req->r_flags & CEPH_OSD_FLAG_WRITE) ||
3611 			     !force_resend_writes))
3612 				break;
3613 
3614 			/* fall through */
3615 		case CALC_TARGET_NEED_RESEND:
3616 			cancel_map_check(req);
3617 			unlink_request(osd, req);
3618 			insert_request(need_resend, req);
3619 			break;
3620 		case CALC_TARGET_POOL_DNE:
3621 			check_pool_dne(req);
3622 			break;
3623 		}
3624 	}
3625 }
3626 
3627 static int handle_one_map(struct ceph_osd_client *osdc,
3628 			  void *p, void *end, bool incremental,
3629 			  struct rb_root *need_resend,
3630 			  struct list_head *need_resend_linger)
3631 {
3632 	struct ceph_osdmap *newmap;
3633 	struct rb_node *n;
3634 	bool skipped_map = false;
3635 	bool was_full;
3636 
3637 	was_full = ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL);
3638 	set_pool_was_full(osdc);
3639 
3640 	if (incremental)
3641 		newmap = osdmap_apply_incremental(&p, end, osdc->osdmap);
3642 	else
3643 		newmap = ceph_osdmap_decode(&p, end);
3644 	if (IS_ERR(newmap))
3645 		return PTR_ERR(newmap);
3646 
3647 	if (newmap != osdc->osdmap) {
3648 		/*
3649 		 * Preserve ->was_full before destroying the old map.
3650 		 * For pools that weren't in the old map, ->was_full
3651 		 * should be false.
3652 		 */
3653 		for (n = rb_first(&newmap->pg_pools); n; n = rb_next(n)) {
3654 			struct ceph_pg_pool_info *pi =
3655 			    rb_entry(n, struct ceph_pg_pool_info, node);
3656 			struct ceph_pg_pool_info *old_pi;
3657 
3658 			old_pi = ceph_pg_pool_by_id(osdc->osdmap, pi->id);
3659 			if (old_pi)
3660 				pi->was_full = old_pi->was_full;
3661 			else
3662 				WARN_ON(pi->was_full);
3663 		}
3664 
3665 		if (osdc->osdmap->epoch &&
3666 		    osdc->osdmap->epoch + 1 < newmap->epoch) {
3667 			WARN_ON(incremental);
3668 			skipped_map = true;
3669 		}
3670 
3671 		ceph_osdmap_destroy(osdc->osdmap);
3672 		osdc->osdmap = newmap;
3673 	}
3674 
3675 	was_full &= !ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL);
3676 	scan_requests(&osdc->homeless_osd, skipped_map, was_full, true,
3677 		      need_resend, need_resend_linger);
3678 
3679 	for (n = rb_first(&osdc->osds); n; ) {
3680 		struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node);
3681 
3682 		n = rb_next(n); /* close_osd() */
3683 
3684 		scan_requests(osd, skipped_map, was_full, true, need_resend,
3685 			      need_resend_linger);
3686 		if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
3687 		    memcmp(&osd->o_con.peer_addr,
3688 			   ceph_osd_addr(osdc->osdmap, osd->o_osd),
3689 			   sizeof(struct ceph_entity_addr)))
3690 			close_osd(osd);
3691 	}
3692 
3693 	return 0;
3694 }
3695 
3696 static void kick_requests(struct ceph_osd_client *osdc,
3697 			  struct rb_root *need_resend,
3698 			  struct list_head *need_resend_linger)
3699 {
3700 	struct ceph_osd_linger_request *lreq, *nlreq;
3701 	enum calc_target_result ct_res;
3702 	struct rb_node *n;
3703 
3704 	/* make sure need_resend targets reflect latest map */
3705 	for (n = rb_first(need_resend); n; ) {
3706 		struct ceph_osd_request *req =
3707 		    rb_entry(n, struct ceph_osd_request, r_node);
3708 
3709 		n = rb_next(n);
3710 
3711 		if (req->r_t.epoch < osdc->osdmap->epoch) {
3712 			ct_res = calc_target(osdc, &req->r_t, NULL, false);
3713 			if (ct_res == CALC_TARGET_POOL_DNE) {
3714 				erase_request(need_resend, req);
3715 				check_pool_dne(req);
3716 			}
3717 		}
3718 	}
3719 
3720 	for (n = rb_first(need_resend); n; ) {
3721 		struct ceph_osd_request *req =
3722 		    rb_entry(n, struct ceph_osd_request, r_node);
3723 		struct ceph_osd *osd;
3724 
3725 		n = rb_next(n);
3726 		erase_request(need_resend, req); /* before link_request() */
3727 
3728 		osd = lookup_create_osd(osdc, req->r_t.osd, true);
3729 		link_request(osd, req);
3730 		if (!req->r_linger) {
3731 			if (!osd_homeless(osd) && !req->r_t.paused)
3732 				send_request(req);
3733 		} else {
3734 			cancel_linger_request(req);
3735 		}
3736 	}
3737 
3738 	list_for_each_entry_safe(lreq, nlreq, need_resend_linger, scan_item) {
3739 		if (!osd_homeless(lreq->osd))
3740 			send_linger(lreq);
3741 
3742 		list_del_init(&lreq->scan_item);
3743 	}
3744 }
3745 
3746 /*
3747  * Process updated osd map.
3748  *
3749  * The message contains any number of incremental and full maps, normally
3750  * indicating some sort of topology change in the cluster.  Kick requests
3751  * off to different OSDs as needed.
3752  */
3753 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
3754 {
3755 	void *p = msg->front.iov_base;
3756 	void *const end = p + msg->front.iov_len;
3757 	u32 nr_maps, maplen;
3758 	u32 epoch;
3759 	struct ceph_fsid fsid;
3760 	struct rb_root need_resend = RB_ROOT;
3761 	LIST_HEAD(need_resend_linger);
3762 	bool handled_incremental = false;
3763 	bool was_pauserd, was_pausewr;
3764 	bool pauserd, pausewr;
3765 	int err;
3766 
3767 	dout("%s have %u\n", __func__, osdc->osdmap->epoch);
3768 	down_write(&osdc->lock);
3769 
3770 	/* verify fsid */
3771 	ceph_decode_need(&p, end, sizeof(fsid), bad);
3772 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
3773 	if (ceph_check_fsid(osdc->client, &fsid) < 0)
3774 		goto bad;
3775 
3776 	was_pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD);
3777 	was_pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) ||
3778 		      ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) ||
3779 		      have_pool_full(osdc);
3780 
3781 	/* incremental maps */
3782 	ceph_decode_32_safe(&p, end, nr_maps, bad);
3783 	dout(" %d inc maps\n", nr_maps);
3784 	while (nr_maps > 0) {
3785 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3786 		epoch = ceph_decode_32(&p);
3787 		maplen = ceph_decode_32(&p);
3788 		ceph_decode_need(&p, end, maplen, bad);
3789 		if (osdc->osdmap->epoch &&
3790 		    osdc->osdmap->epoch + 1 == epoch) {
3791 			dout("applying incremental map %u len %d\n",
3792 			     epoch, maplen);
3793 			err = handle_one_map(osdc, p, p + maplen, true,
3794 					     &need_resend, &need_resend_linger);
3795 			if (err)
3796 				goto bad;
3797 			handled_incremental = true;
3798 		} else {
3799 			dout("ignoring incremental map %u len %d\n",
3800 			     epoch, maplen);
3801 		}
3802 		p += maplen;
3803 		nr_maps--;
3804 	}
3805 	if (handled_incremental)
3806 		goto done;
3807 
3808 	/* full maps */
3809 	ceph_decode_32_safe(&p, end, nr_maps, bad);
3810 	dout(" %d full maps\n", nr_maps);
3811 	while (nr_maps) {
3812 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3813 		epoch = ceph_decode_32(&p);
3814 		maplen = ceph_decode_32(&p);
3815 		ceph_decode_need(&p, end, maplen, bad);
3816 		if (nr_maps > 1) {
3817 			dout("skipping non-latest full map %u len %d\n",
3818 			     epoch, maplen);
3819 		} else if (osdc->osdmap->epoch >= epoch) {
3820 			dout("skipping full map %u len %d, "
3821 			     "older than our %u\n", epoch, maplen,
3822 			     osdc->osdmap->epoch);
3823 		} else {
3824 			dout("taking full map %u len %d\n", epoch, maplen);
3825 			err = handle_one_map(osdc, p, p + maplen, false,
3826 					     &need_resend, &need_resend_linger);
3827 			if (err)
3828 				goto bad;
3829 		}
3830 		p += maplen;
3831 		nr_maps--;
3832 	}
3833 
3834 done:
3835 	/*
3836 	 * subscribe to subsequent osdmap updates if full to ensure
3837 	 * we find out when we are no longer full and stop returning
3838 	 * ENOSPC.
3839 	 */
3840 	pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD);
3841 	pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) ||
3842 		  ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) ||
3843 		  have_pool_full(osdc);
3844 	if (was_pauserd || was_pausewr || pauserd || pausewr ||
3845 	    osdc->osdmap->epoch < osdc->epoch_barrier)
3846 		maybe_request_map(osdc);
3847 
3848 	kick_requests(osdc, &need_resend, &need_resend_linger);
3849 
3850 	ceph_osdc_abort_on_full(osdc);
3851 	ceph_monc_got_map(&osdc->client->monc, CEPH_SUB_OSDMAP,
3852 			  osdc->osdmap->epoch);
3853 	up_write(&osdc->lock);
3854 	wake_up_all(&osdc->client->auth_wq);
3855 	return;
3856 
3857 bad:
3858 	pr_err("osdc handle_map corrupt msg\n");
3859 	ceph_msg_dump(msg);
3860 	up_write(&osdc->lock);
3861 }
3862 
3863 /*
3864  * Resubmit requests pending on the given osd.
3865  */
3866 static void kick_osd_requests(struct ceph_osd *osd)
3867 {
3868 	struct rb_node *n;
3869 
3870 	clear_backoffs(osd);
3871 
3872 	for (n = rb_first(&osd->o_requests); n; ) {
3873 		struct ceph_osd_request *req =
3874 		    rb_entry(n, struct ceph_osd_request, r_node);
3875 
3876 		n = rb_next(n); /* cancel_linger_request() */
3877 
3878 		if (!req->r_linger) {
3879 			if (!req->r_t.paused)
3880 				send_request(req);
3881 		} else {
3882 			cancel_linger_request(req);
3883 		}
3884 	}
3885 	for (n = rb_first(&osd->o_linger_requests); n; n = rb_next(n)) {
3886 		struct ceph_osd_linger_request *lreq =
3887 		    rb_entry(n, struct ceph_osd_linger_request, node);
3888 
3889 		send_linger(lreq);
3890 	}
3891 }
3892 
3893 /*
3894  * If the osd connection drops, we need to resubmit all requests.
3895  */
3896 static void osd_fault(struct ceph_connection *con)
3897 {
3898 	struct ceph_osd *osd = con->private;
3899 	struct ceph_osd_client *osdc = osd->o_osdc;
3900 
3901 	dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd);
3902 
3903 	down_write(&osdc->lock);
3904 	if (!osd_registered(osd)) {
3905 		dout("%s osd%d unknown\n", __func__, osd->o_osd);
3906 		goto out_unlock;
3907 	}
3908 
3909 	if (!reopen_osd(osd))
3910 		kick_osd_requests(osd);
3911 	maybe_request_map(osdc);
3912 
3913 out_unlock:
3914 	up_write(&osdc->lock);
3915 }
3916 
3917 struct MOSDBackoff {
3918 	struct ceph_spg spgid;
3919 	u32 map_epoch;
3920 	u8 op;
3921 	u64 id;
3922 	struct ceph_hobject_id *begin;
3923 	struct ceph_hobject_id *end;
3924 };
3925 
3926 static int decode_MOSDBackoff(const struct ceph_msg *msg, struct MOSDBackoff *m)
3927 {
3928 	void *p = msg->front.iov_base;
3929 	void *const end = p + msg->front.iov_len;
3930 	u8 struct_v;
3931 	u32 struct_len;
3932 	int ret;
3933 
3934 	ret = ceph_start_decoding(&p, end, 1, "spg_t", &struct_v, &struct_len);
3935 	if (ret)
3936 		return ret;
3937 
3938 	ret = ceph_decode_pgid(&p, end, &m->spgid.pgid);
3939 	if (ret)
3940 		return ret;
3941 
3942 	ceph_decode_8_safe(&p, end, m->spgid.shard, e_inval);
3943 	ceph_decode_32_safe(&p, end, m->map_epoch, e_inval);
3944 	ceph_decode_8_safe(&p, end, m->op, e_inval);
3945 	ceph_decode_64_safe(&p, end, m->id, e_inval);
3946 
3947 	m->begin = kzalloc(sizeof(*m->begin), GFP_NOIO);
3948 	if (!m->begin)
3949 		return -ENOMEM;
3950 
3951 	ret = decode_hoid(&p, end, m->begin);
3952 	if (ret) {
3953 		free_hoid(m->begin);
3954 		return ret;
3955 	}
3956 
3957 	m->end = kzalloc(sizeof(*m->end), GFP_NOIO);
3958 	if (!m->end) {
3959 		free_hoid(m->begin);
3960 		return -ENOMEM;
3961 	}
3962 
3963 	ret = decode_hoid(&p, end, m->end);
3964 	if (ret) {
3965 		free_hoid(m->begin);
3966 		free_hoid(m->end);
3967 		return ret;
3968 	}
3969 
3970 	return 0;
3971 
3972 e_inval:
3973 	return -EINVAL;
3974 }
3975 
3976 static struct ceph_msg *create_backoff_message(
3977 				const struct ceph_osd_backoff *backoff,
3978 				u32 map_epoch)
3979 {
3980 	struct ceph_msg *msg;
3981 	void *p, *end;
3982 	int msg_size;
3983 
3984 	msg_size = CEPH_ENCODING_START_BLK_LEN +
3985 			CEPH_PGID_ENCODING_LEN + 1; /* spgid */
3986 	msg_size += 4 + 1 + 8; /* map_epoch, op, id */
3987 	msg_size += CEPH_ENCODING_START_BLK_LEN +
3988 			hoid_encoding_size(backoff->begin);
3989 	msg_size += CEPH_ENCODING_START_BLK_LEN +
3990 			hoid_encoding_size(backoff->end);
3991 
3992 	msg = ceph_msg_new(CEPH_MSG_OSD_BACKOFF, msg_size, GFP_NOIO, true);
3993 	if (!msg)
3994 		return NULL;
3995 
3996 	p = msg->front.iov_base;
3997 	end = p + msg->front_alloc_len;
3998 
3999 	encode_spgid(&p, &backoff->spgid);
4000 	ceph_encode_32(&p, map_epoch);
4001 	ceph_encode_8(&p, CEPH_OSD_BACKOFF_OP_ACK_BLOCK);
4002 	ceph_encode_64(&p, backoff->id);
4003 	encode_hoid(&p, end, backoff->begin);
4004 	encode_hoid(&p, end, backoff->end);
4005 	BUG_ON(p != end);
4006 
4007 	msg->front.iov_len = p - msg->front.iov_base;
4008 	msg->hdr.version = cpu_to_le16(1); /* MOSDBackoff v1 */
4009 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
4010 
4011 	return msg;
4012 }
4013 
4014 static void handle_backoff_block(struct ceph_osd *osd, struct MOSDBackoff *m)
4015 {
4016 	struct ceph_spg_mapping *spg;
4017 	struct ceph_osd_backoff *backoff;
4018 	struct ceph_msg *msg;
4019 
4020 	dout("%s osd%d spgid %llu.%xs%d id %llu\n", __func__, osd->o_osd,
4021 	     m->spgid.pgid.pool, m->spgid.pgid.seed, m->spgid.shard, m->id);
4022 
4023 	spg = lookup_spg_mapping(&osd->o_backoff_mappings, &m->spgid);
4024 	if (!spg) {
4025 		spg = alloc_spg_mapping();
4026 		if (!spg) {
4027 			pr_err("%s failed to allocate spg\n", __func__);
4028 			return;
4029 		}
4030 		spg->spgid = m->spgid; /* struct */
4031 		insert_spg_mapping(&osd->o_backoff_mappings, spg);
4032 	}
4033 
4034 	backoff = alloc_backoff();
4035 	if (!backoff) {
4036 		pr_err("%s failed to allocate backoff\n", __func__);
4037 		return;
4038 	}
4039 	backoff->spgid = m->spgid; /* struct */
4040 	backoff->id = m->id;
4041 	backoff->begin = m->begin;
4042 	m->begin = NULL; /* backoff now owns this */
4043 	backoff->end = m->end;
4044 	m->end = NULL;   /* ditto */
4045 
4046 	insert_backoff(&spg->backoffs, backoff);
4047 	insert_backoff_by_id(&osd->o_backoffs_by_id, backoff);
4048 
4049 	/*
4050 	 * Ack with original backoff's epoch so that the OSD can
4051 	 * discard this if there was a PG split.
4052 	 */
4053 	msg = create_backoff_message(backoff, m->map_epoch);
4054 	if (!msg) {
4055 		pr_err("%s failed to allocate msg\n", __func__);
4056 		return;
4057 	}
4058 	ceph_con_send(&osd->o_con, msg);
4059 }
4060 
4061 static bool target_contained_by(const struct ceph_osd_request_target *t,
4062 				const struct ceph_hobject_id *begin,
4063 				const struct ceph_hobject_id *end)
4064 {
4065 	struct ceph_hobject_id hoid;
4066 	int cmp;
4067 
4068 	hoid_fill_from_target(&hoid, t);
4069 	cmp = hoid_compare(&hoid, begin);
4070 	return !cmp || (cmp > 0 && hoid_compare(&hoid, end) < 0);
4071 }
4072 
4073 static void handle_backoff_unblock(struct ceph_osd *osd,
4074 				   const struct MOSDBackoff *m)
4075 {
4076 	struct ceph_spg_mapping *spg;
4077 	struct ceph_osd_backoff *backoff;
4078 	struct rb_node *n;
4079 
4080 	dout("%s osd%d spgid %llu.%xs%d id %llu\n", __func__, osd->o_osd,
4081 	     m->spgid.pgid.pool, m->spgid.pgid.seed, m->spgid.shard, m->id);
4082 
4083 	backoff = lookup_backoff_by_id(&osd->o_backoffs_by_id, m->id);
4084 	if (!backoff) {
4085 		pr_err("%s osd%d spgid %llu.%xs%d id %llu backoff dne\n",
4086 		       __func__, osd->o_osd, m->spgid.pgid.pool,
4087 		       m->spgid.pgid.seed, m->spgid.shard, m->id);
4088 		return;
4089 	}
4090 
4091 	if (hoid_compare(backoff->begin, m->begin) &&
4092 	    hoid_compare(backoff->end, m->end)) {
4093 		pr_err("%s osd%d spgid %llu.%xs%d id %llu bad range?\n",
4094 		       __func__, osd->o_osd, m->spgid.pgid.pool,
4095 		       m->spgid.pgid.seed, m->spgid.shard, m->id);
4096 		/* unblock it anyway... */
4097 	}
4098 
4099 	spg = lookup_spg_mapping(&osd->o_backoff_mappings, &backoff->spgid);
4100 	BUG_ON(!spg);
4101 
4102 	erase_backoff(&spg->backoffs, backoff);
4103 	erase_backoff_by_id(&osd->o_backoffs_by_id, backoff);
4104 	free_backoff(backoff);
4105 
4106 	if (RB_EMPTY_ROOT(&spg->backoffs)) {
4107 		erase_spg_mapping(&osd->o_backoff_mappings, spg);
4108 		free_spg_mapping(spg);
4109 	}
4110 
4111 	for (n = rb_first(&osd->o_requests); n; n = rb_next(n)) {
4112 		struct ceph_osd_request *req =
4113 		    rb_entry(n, struct ceph_osd_request, r_node);
4114 
4115 		if (!ceph_spg_compare(&req->r_t.spgid, &m->spgid)) {
4116 			/*
4117 			 * Match against @m, not @backoff -- the PG may
4118 			 * have split on the OSD.
4119 			 */
4120 			if (target_contained_by(&req->r_t, m->begin, m->end)) {
4121 				/*
4122 				 * If no other installed backoff applies,
4123 				 * resend.
4124 				 */
4125 				send_request(req);
4126 			}
4127 		}
4128 	}
4129 }
4130 
4131 static void handle_backoff(struct ceph_osd *osd, struct ceph_msg *msg)
4132 {
4133 	struct ceph_osd_client *osdc = osd->o_osdc;
4134 	struct MOSDBackoff m;
4135 	int ret;
4136 
4137 	down_read(&osdc->lock);
4138 	if (!osd_registered(osd)) {
4139 		dout("%s osd%d unknown\n", __func__, osd->o_osd);
4140 		up_read(&osdc->lock);
4141 		return;
4142 	}
4143 	WARN_ON(osd->o_osd != le64_to_cpu(msg->hdr.src.num));
4144 
4145 	mutex_lock(&osd->lock);
4146 	ret = decode_MOSDBackoff(msg, &m);
4147 	if (ret) {
4148 		pr_err("failed to decode MOSDBackoff: %d\n", ret);
4149 		ceph_msg_dump(msg);
4150 		goto out_unlock;
4151 	}
4152 
4153 	switch (m.op) {
4154 	case CEPH_OSD_BACKOFF_OP_BLOCK:
4155 		handle_backoff_block(osd, &m);
4156 		break;
4157 	case CEPH_OSD_BACKOFF_OP_UNBLOCK:
4158 		handle_backoff_unblock(osd, &m);
4159 		break;
4160 	default:
4161 		pr_err("%s osd%d unknown op %d\n", __func__, osd->o_osd, m.op);
4162 	}
4163 
4164 	free_hoid(m.begin);
4165 	free_hoid(m.end);
4166 
4167 out_unlock:
4168 	mutex_unlock(&osd->lock);
4169 	up_read(&osdc->lock);
4170 }
4171 
4172 /*
4173  * Process osd watch notifications
4174  */
4175 static void handle_watch_notify(struct ceph_osd_client *osdc,
4176 				struct ceph_msg *msg)
4177 {
4178 	void *p = msg->front.iov_base;
4179 	void *const end = p + msg->front.iov_len;
4180 	struct ceph_osd_linger_request *lreq;
4181 	struct linger_work *lwork;
4182 	u8 proto_ver, opcode;
4183 	u64 cookie, notify_id;
4184 	u64 notifier_id = 0;
4185 	s32 return_code = 0;
4186 	void *payload = NULL;
4187 	u32 payload_len = 0;
4188 
4189 	ceph_decode_8_safe(&p, end, proto_ver, bad);
4190 	ceph_decode_8_safe(&p, end, opcode, bad);
4191 	ceph_decode_64_safe(&p, end, cookie, bad);
4192 	p += 8; /* skip ver */
4193 	ceph_decode_64_safe(&p, end, notify_id, bad);
4194 
4195 	if (proto_ver >= 1) {
4196 		ceph_decode_32_safe(&p, end, payload_len, bad);
4197 		ceph_decode_need(&p, end, payload_len, bad);
4198 		payload = p;
4199 		p += payload_len;
4200 	}
4201 
4202 	if (le16_to_cpu(msg->hdr.version) >= 2)
4203 		ceph_decode_32_safe(&p, end, return_code, bad);
4204 
4205 	if (le16_to_cpu(msg->hdr.version) >= 3)
4206 		ceph_decode_64_safe(&p, end, notifier_id, bad);
4207 
4208 	down_read(&osdc->lock);
4209 	lreq = lookup_linger_osdc(&osdc->linger_requests, cookie);
4210 	if (!lreq) {
4211 		dout("%s opcode %d cookie %llu dne\n", __func__, opcode,
4212 		     cookie);
4213 		goto out_unlock_osdc;
4214 	}
4215 
4216 	mutex_lock(&lreq->lock);
4217 	dout("%s opcode %d cookie %llu lreq %p is_watch %d\n", __func__,
4218 	     opcode, cookie, lreq, lreq->is_watch);
4219 	if (opcode == CEPH_WATCH_EVENT_DISCONNECT) {
4220 		if (!lreq->last_error) {
4221 			lreq->last_error = -ENOTCONN;
4222 			queue_watch_error(lreq);
4223 		}
4224 	} else if (!lreq->is_watch) {
4225 		/* CEPH_WATCH_EVENT_NOTIFY_COMPLETE */
4226 		if (lreq->notify_id && lreq->notify_id != notify_id) {
4227 			dout("lreq %p notify_id %llu != %llu, ignoring\n", lreq,
4228 			     lreq->notify_id, notify_id);
4229 		} else if (!completion_done(&lreq->notify_finish_wait)) {
4230 			struct ceph_msg_data *data =
4231 			    list_first_entry_or_null(&msg->data,
4232 						     struct ceph_msg_data,
4233 						     links);
4234 
4235 			if (data) {
4236 				if (lreq->preply_pages) {
4237 					WARN_ON(data->type !=
4238 							CEPH_MSG_DATA_PAGES);
4239 					*lreq->preply_pages = data->pages;
4240 					*lreq->preply_len = data->length;
4241 				} else {
4242 					ceph_release_page_vector(data->pages,
4243 					       calc_pages_for(0, data->length));
4244 				}
4245 			}
4246 			lreq->notify_finish_error = return_code;
4247 			complete_all(&lreq->notify_finish_wait);
4248 		}
4249 	} else {
4250 		/* CEPH_WATCH_EVENT_NOTIFY */
4251 		lwork = lwork_alloc(lreq, do_watch_notify);
4252 		if (!lwork) {
4253 			pr_err("failed to allocate notify-lwork\n");
4254 			goto out_unlock_lreq;
4255 		}
4256 
4257 		lwork->notify.notify_id = notify_id;
4258 		lwork->notify.notifier_id = notifier_id;
4259 		lwork->notify.payload = payload;
4260 		lwork->notify.payload_len = payload_len;
4261 		lwork->notify.msg = ceph_msg_get(msg);
4262 		lwork_queue(lwork);
4263 	}
4264 
4265 out_unlock_lreq:
4266 	mutex_unlock(&lreq->lock);
4267 out_unlock_osdc:
4268 	up_read(&osdc->lock);
4269 	return;
4270 
4271 bad:
4272 	pr_err("osdc handle_watch_notify corrupt msg\n");
4273 }
4274 
4275 /*
4276  * Register request, send initial attempt.
4277  */
4278 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
4279 			    struct ceph_osd_request *req,
4280 			    bool nofail)
4281 {
4282 	down_read(&osdc->lock);
4283 	submit_request(req, false);
4284 	up_read(&osdc->lock);
4285 
4286 	return 0;
4287 }
4288 EXPORT_SYMBOL(ceph_osdc_start_request);
4289 
4290 /*
4291  * Unregister a registered request.  The request is not completed:
4292  * ->r_result isn't set and __complete_request() isn't called.
4293  */
4294 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
4295 {
4296 	struct ceph_osd_client *osdc = req->r_osdc;
4297 
4298 	down_write(&osdc->lock);
4299 	if (req->r_osd)
4300 		cancel_request(req);
4301 	up_write(&osdc->lock);
4302 }
4303 EXPORT_SYMBOL(ceph_osdc_cancel_request);
4304 
4305 /*
4306  * @timeout: in jiffies, 0 means "wait forever"
4307  */
4308 static int wait_request_timeout(struct ceph_osd_request *req,
4309 				unsigned long timeout)
4310 {
4311 	long left;
4312 
4313 	dout("%s req %p tid %llu\n", __func__, req, req->r_tid);
4314 	left = wait_for_completion_killable_timeout(&req->r_completion,
4315 						ceph_timeout_jiffies(timeout));
4316 	if (left <= 0) {
4317 		left = left ?: -ETIMEDOUT;
4318 		ceph_osdc_cancel_request(req);
4319 	} else {
4320 		left = req->r_result; /* completed */
4321 	}
4322 
4323 	return left;
4324 }
4325 
4326 /*
4327  * wait for a request to complete
4328  */
4329 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
4330 			   struct ceph_osd_request *req)
4331 {
4332 	return wait_request_timeout(req, 0);
4333 }
4334 EXPORT_SYMBOL(ceph_osdc_wait_request);
4335 
4336 /*
4337  * sync - wait for all in-flight requests to flush.  avoid starvation.
4338  */
4339 void ceph_osdc_sync(struct ceph_osd_client *osdc)
4340 {
4341 	struct rb_node *n, *p;
4342 	u64 last_tid = atomic64_read(&osdc->last_tid);
4343 
4344 again:
4345 	down_read(&osdc->lock);
4346 	for (n = rb_first(&osdc->osds); n; n = rb_next(n)) {
4347 		struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node);
4348 
4349 		mutex_lock(&osd->lock);
4350 		for (p = rb_first(&osd->o_requests); p; p = rb_next(p)) {
4351 			struct ceph_osd_request *req =
4352 			    rb_entry(p, struct ceph_osd_request, r_node);
4353 
4354 			if (req->r_tid > last_tid)
4355 				break;
4356 
4357 			if (!(req->r_flags & CEPH_OSD_FLAG_WRITE))
4358 				continue;
4359 
4360 			ceph_osdc_get_request(req);
4361 			mutex_unlock(&osd->lock);
4362 			up_read(&osdc->lock);
4363 			dout("%s waiting on req %p tid %llu last_tid %llu\n",
4364 			     __func__, req, req->r_tid, last_tid);
4365 			wait_for_completion(&req->r_completion);
4366 			ceph_osdc_put_request(req);
4367 			goto again;
4368 		}
4369 
4370 		mutex_unlock(&osd->lock);
4371 	}
4372 
4373 	up_read(&osdc->lock);
4374 	dout("%s done last_tid %llu\n", __func__, last_tid);
4375 }
4376 EXPORT_SYMBOL(ceph_osdc_sync);
4377 
4378 static struct ceph_osd_request *
4379 alloc_linger_request(struct ceph_osd_linger_request *lreq)
4380 {
4381 	struct ceph_osd_request *req;
4382 
4383 	req = ceph_osdc_alloc_request(lreq->osdc, NULL, 1, false, GFP_NOIO);
4384 	if (!req)
4385 		return NULL;
4386 
4387 	ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid);
4388 	ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc);
4389 
4390 	if (ceph_osdc_alloc_messages(req, GFP_NOIO)) {
4391 		ceph_osdc_put_request(req);
4392 		return NULL;
4393 	}
4394 
4395 	return req;
4396 }
4397 
4398 /*
4399  * Returns a handle, caller owns a ref.
4400  */
4401 struct ceph_osd_linger_request *
4402 ceph_osdc_watch(struct ceph_osd_client *osdc,
4403 		struct ceph_object_id *oid,
4404 		struct ceph_object_locator *oloc,
4405 		rados_watchcb2_t wcb,
4406 		rados_watcherrcb_t errcb,
4407 		void *data)
4408 {
4409 	struct ceph_osd_linger_request *lreq;
4410 	int ret;
4411 
4412 	lreq = linger_alloc(osdc);
4413 	if (!lreq)
4414 		return ERR_PTR(-ENOMEM);
4415 
4416 	lreq->is_watch = true;
4417 	lreq->wcb = wcb;
4418 	lreq->errcb = errcb;
4419 	lreq->data = data;
4420 	lreq->watch_valid_thru = jiffies;
4421 
4422 	ceph_oid_copy(&lreq->t.base_oid, oid);
4423 	ceph_oloc_copy(&lreq->t.base_oloc, oloc);
4424 	lreq->t.flags = CEPH_OSD_FLAG_WRITE;
4425 	ktime_get_real_ts(&lreq->mtime);
4426 
4427 	lreq->reg_req = alloc_linger_request(lreq);
4428 	if (!lreq->reg_req) {
4429 		ret = -ENOMEM;
4430 		goto err_put_lreq;
4431 	}
4432 
4433 	lreq->ping_req = alloc_linger_request(lreq);
4434 	if (!lreq->ping_req) {
4435 		ret = -ENOMEM;
4436 		goto err_put_lreq;
4437 	}
4438 
4439 	down_write(&osdc->lock);
4440 	linger_register(lreq); /* before osd_req_op_* */
4441 	osd_req_op_watch_init(lreq->reg_req, 0, lreq->linger_id,
4442 			      CEPH_OSD_WATCH_OP_WATCH);
4443 	osd_req_op_watch_init(lreq->ping_req, 0, lreq->linger_id,
4444 			      CEPH_OSD_WATCH_OP_PING);
4445 	linger_submit(lreq);
4446 	up_write(&osdc->lock);
4447 
4448 	ret = linger_reg_commit_wait(lreq);
4449 	if (ret) {
4450 		linger_cancel(lreq);
4451 		goto err_put_lreq;
4452 	}
4453 
4454 	return lreq;
4455 
4456 err_put_lreq:
4457 	linger_put(lreq);
4458 	return ERR_PTR(ret);
4459 }
4460 EXPORT_SYMBOL(ceph_osdc_watch);
4461 
4462 /*
4463  * Releases a ref.
4464  *
4465  * Times out after mount_timeout to preserve rbd unmap behaviour
4466  * introduced in 2894e1d76974 ("rbd: timeout watch teardown on unmap
4467  * with mount_timeout").
4468  */
4469 int ceph_osdc_unwatch(struct ceph_osd_client *osdc,
4470 		      struct ceph_osd_linger_request *lreq)
4471 {
4472 	struct ceph_options *opts = osdc->client->options;
4473 	struct ceph_osd_request *req;
4474 	int ret;
4475 
4476 	req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO);
4477 	if (!req)
4478 		return -ENOMEM;
4479 
4480 	ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid);
4481 	ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc);
4482 	req->r_flags = CEPH_OSD_FLAG_WRITE;
4483 	ktime_get_real_ts(&req->r_mtime);
4484 	osd_req_op_watch_init(req, 0, lreq->linger_id,
4485 			      CEPH_OSD_WATCH_OP_UNWATCH);
4486 
4487 	ret = ceph_osdc_alloc_messages(req, GFP_NOIO);
4488 	if (ret)
4489 		goto out_put_req;
4490 
4491 	ceph_osdc_start_request(osdc, req, false);
4492 	linger_cancel(lreq);
4493 	linger_put(lreq);
4494 	ret = wait_request_timeout(req, opts->mount_timeout);
4495 
4496 out_put_req:
4497 	ceph_osdc_put_request(req);
4498 	return ret;
4499 }
4500 EXPORT_SYMBOL(ceph_osdc_unwatch);
4501 
4502 static int osd_req_op_notify_ack_init(struct ceph_osd_request *req, int which,
4503 				      u64 notify_id, u64 cookie, void *payload,
4504 				      size_t payload_len)
4505 {
4506 	struct ceph_osd_req_op *op;
4507 	struct ceph_pagelist *pl;
4508 	int ret;
4509 
4510 	op = _osd_req_op_init(req, which, CEPH_OSD_OP_NOTIFY_ACK, 0);
4511 
4512 	pl = kmalloc(sizeof(*pl), GFP_NOIO);
4513 	if (!pl)
4514 		return -ENOMEM;
4515 
4516 	ceph_pagelist_init(pl);
4517 	ret = ceph_pagelist_encode_64(pl, notify_id);
4518 	ret |= ceph_pagelist_encode_64(pl, cookie);
4519 	if (payload) {
4520 		ret |= ceph_pagelist_encode_32(pl, payload_len);
4521 		ret |= ceph_pagelist_append(pl, payload, payload_len);
4522 	} else {
4523 		ret |= ceph_pagelist_encode_32(pl, 0);
4524 	}
4525 	if (ret) {
4526 		ceph_pagelist_release(pl);
4527 		return -ENOMEM;
4528 	}
4529 
4530 	ceph_osd_data_pagelist_init(&op->notify_ack.request_data, pl);
4531 	op->indata_len = pl->length;
4532 	return 0;
4533 }
4534 
4535 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc,
4536 			 struct ceph_object_id *oid,
4537 			 struct ceph_object_locator *oloc,
4538 			 u64 notify_id,
4539 			 u64 cookie,
4540 			 void *payload,
4541 			 size_t payload_len)
4542 {
4543 	struct ceph_osd_request *req;
4544 	int ret;
4545 
4546 	req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO);
4547 	if (!req)
4548 		return -ENOMEM;
4549 
4550 	ceph_oid_copy(&req->r_base_oid, oid);
4551 	ceph_oloc_copy(&req->r_base_oloc, oloc);
4552 	req->r_flags = CEPH_OSD_FLAG_READ;
4553 
4554 	ret = ceph_osdc_alloc_messages(req, GFP_NOIO);
4555 	if (ret)
4556 		goto out_put_req;
4557 
4558 	ret = osd_req_op_notify_ack_init(req, 0, notify_id, cookie, payload,
4559 					 payload_len);
4560 	if (ret)
4561 		goto out_put_req;
4562 
4563 	ceph_osdc_start_request(osdc, req, false);
4564 	ret = ceph_osdc_wait_request(osdc, req);
4565 
4566 out_put_req:
4567 	ceph_osdc_put_request(req);
4568 	return ret;
4569 }
4570 EXPORT_SYMBOL(ceph_osdc_notify_ack);
4571 
4572 static int osd_req_op_notify_init(struct ceph_osd_request *req, int which,
4573 				  u64 cookie, u32 prot_ver, u32 timeout,
4574 				  void *payload, size_t payload_len)
4575 {
4576 	struct ceph_osd_req_op *op;
4577 	struct ceph_pagelist *pl;
4578 	int ret;
4579 
4580 	op = _osd_req_op_init(req, which, CEPH_OSD_OP_NOTIFY, 0);
4581 	op->notify.cookie = cookie;
4582 
4583 	pl = kmalloc(sizeof(*pl), GFP_NOIO);
4584 	if (!pl)
4585 		return -ENOMEM;
4586 
4587 	ceph_pagelist_init(pl);
4588 	ret = ceph_pagelist_encode_32(pl, 1); /* prot_ver */
4589 	ret |= ceph_pagelist_encode_32(pl, timeout);
4590 	ret |= ceph_pagelist_encode_32(pl, payload_len);
4591 	ret |= ceph_pagelist_append(pl, payload, payload_len);
4592 	if (ret) {
4593 		ceph_pagelist_release(pl);
4594 		return -ENOMEM;
4595 	}
4596 
4597 	ceph_osd_data_pagelist_init(&op->notify.request_data, pl);
4598 	op->indata_len = pl->length;
4599 	return 0;
4600 }
4601 
4602 /*
4603  * @timeout: in seconds
4604  *
4605  * @preply_{pages,len} are initialized both on success and error.
4606  * The caller is responsible for:
4607  *
4608  *     ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len))
4609  */
4610 int ceph_osdc_notify(struct ceph_osd_client *osdc,
4611 		     struct ceph_object_id *oid,
4612 		     struct ceph_object_locator *oloc,
4613 		     void *payload,
4614 		     size_t payload_len,
4615 		     u32 timeout,
4616 		     struct page ***preply_pages,
4617 		     size_t *preply_len)
4618 {
4619 	struct ceph_osd_linger_request *lreq;
4620 	struct page **pages;
4621 	int ret;
4622 
4623 	WARN_ON(!timeout);
4624 	if (preply_pages) {
4625 		*preply_pages = NULL;
4626 		*preply_len = 0;
4627 	}
4628 
4629 	lreq = linger_alloc(osdc);
4630 	if (!lreq)
4631 		return -ENOMEM;
4632 
4633 	lreq->preply_pages = preply_pages;
4634 	lreq->preply_len = preply_len;
4635 
4636 	ceph_oid_copy(&lreq->t.base_oid, oid);
4637 	ceph_oloc_copy(&lreq->t.base_oloc, oloc);
4638 	lreq->t.flags = CEPH_OSD_FLAG_READ;
4639 
4640 	lreq->reg_req = alloc_linger_request(lreq);
4641 	if (!lreq->reg_req) {
4642 		ret = -ENOMEM;
4643 		goto out_put_lreq;
4644 	}
4645 
4646 	/* for notify_id */
4647 	pages = ceph_alloc_page_vector(1, GFP_NOIO);
4648 	if (IS_ERR(pages)) {
4649 		ret = PTR_ERR(pages);
4650 		goto out_put_lreq;
4651 	}
4652 
4653 	down_write(&osdc->lock);
4654 	linger_register(lreq); /* before osd_req_op_* */
4655 	ret = osd_req_op_notify_init(lreq->reg_req, 0, lreq->linger_id, 1,
4656 				     timeout, payload, payload_len);
4657 	if (ret) {
4658 		linger_unregister(lreq);
4659 		up_write(&osdc->lock);
4660 		ceph_release_page_vector(pages, 1);
4661 		goto out_put_lreq;
4662 	}
4663 	ceph_osd_data_pages_init(osd_req_op_data(lreq->reg_req, 0, notify,
4664 						 response_data),
4665 				 pages, PAGE_SIZE, 0, false, true);
4666 	linger_submit(lreq);
4667 	up_write(&osdc->lock);
4668 
4669 	ret = linger_reg_commit_wait(lreq);
4670 	if (!ret)
4671 		ret = linger_notify_finish_wait(lreq);
4672 	else
4673 		dout("lreq %p failed to initiate notify %d\n", lreq, ret);
4674 
4675 	linger_cancel(lreq);
4676 out_put_lreq:
4677 	linger_put(lreq);
4678 	return ret;
4679 }
4680 EXPORT_SYMBOL(ceph_osdc_notify);
4681 
4682 /*
4683  * Return the number of milliseconds since the watch was last
4684  * confirmed, or an error.  If there is an error, the watch is no
4685  * longer valid, and should be destroyed with ceph_osdc_unwatch().
4686  */
4687 int ceph_osdc_watch_check(struct ceph_osd_client *osdc,
4688 			  struct ceph_osd_linger_request *lreq)
4689 {
4690 	unsigned long stamp, age;
4691 	int ret;
4692 
4693 	down_read(&osdc->lock);
4694 	mutex_lock(&lreq->lock);
4695 	stamp = lreq->watch_valid_thru;
4696 	if (!list_empty(&lreq->pending_lworks)) {
4697 		struct linger_work *lwork =
4698 		    list_first_entry(&lreq->pending_lworks,
4699 				     struct linger_work,
4700 				     pending_item);
4701 
4702 		if (time_before(lwork->queued_stamp, stamp))
4703 			stamp = lwork->queued_stamp;
4704 	}
4705 	age = jiffies - stamp;
4706 	dout("%s lreq %p linger_id %llu age %lu last_error %d\n", __func__,
4707 	     lreq, lreq->linger_id, age, lreq->last_error);
4708 	/* we are truncating to msecs, so return a safe upper bound */
4709 	ret = lreq->last_error ?: 1 + jiffies_to_msecs(age);
4710 
4711 	mutex_unlock(&lreq->lock);
4712 	up_read(&osdc->lock);
4713 	return ret;
4714 }
4715 
4716 static int decode_watcher(void **p, void *end, struct ceph_watch_item *item)
4717 {
4718 	u8 struct_v;
4719 	u32 struct_len;
4720 	int ret;
4721 
4722 	ret = ceph_start_decoding(p, end, 2, "watch_item_t",
4723 				  &struct_v, &struct_len);
4724 	if (ret)
4725 		return ret;
4726 
4727 	ceph_decode_copy(p, &item->name, sizeof(item->name));
4728 	item->cookie = ceph_decode_64(p);
4729 	*p += 4; /* skip timeout_seconds */
4730 	if (struct_v >= 2) {
4731 		ceph_decode_copy(p, &item->addr, sizeof(item->addr));
4732 		ceph_decode_addr(&item->addr);
4733 	}
4734 
4735 	dout("%s %s%llu cookie %llu addr %s\n", __func__,
4736 	     ENTITY_NAME(item->name), item->cookie,
4737 	     ceph_pr_addr(&item->addr.in_addr));
4738 	return 0;
4739 }
4740 
4741 static int decode_watchers(void **p, void *end,
4742 			   struct ceph_watch_item **watchers,
4743 			   u32 *num_watchers)
4744 {
4745 	u8 struct_v;
4746 	u32 struct_len;
4747 	int i;
4748 	int ret;
4749 
4750 	ret = ceph_start_decoding(p, end, 1, "obj_list_watch_response_t",
4751 				  &struct_v, &struct_len);
4752 	if (ret)
4753 		return ret;
4754 
4755 	*num_watchers = ceph_decode_32(p);
4756 	*watchers = kcalloc(*num_watchers, sizeof(**watchers), GFP_NOIO);
4757 	if (!*watchers)
4758 		return -ENOMEM;
4759 
4760 	for (i = 0; i < *num_watchers; i++) {
4761 		ret = decode_watcher(p, end, *watchers + i);
4762 		if (ret) {
4763 			kfree(*watchers);
4764 			return ret;
4765 		}
4766 	}
4767 
4768 	return 0;
4769 }
4770 
4771 /*
4772  * On success, the caller is responsible for:
4773  *
4774  *     kfree(watchers);
4775  */
4776 int ceph_osdc_list_watchers(struct ceph_osd_client *osdc,
4777 			    struct ceph_object_id *oid,
4778 			    struct ceph_object_locator *oloc,
4779 			    struct ceph_watch_item **watchers,
4780 			    u32 *num_watchers)
4781 {
4782 	struct ceph_osd_request *req;
4783 	struct page **pages;
4784 	int ret;
4785 
4786 	req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO);
4787 	if (!req)
4788 		return -ENOMEM;
4789 
4790 	ceph_oid_copy(&req->r_base_oid, oid);
4791 	ceph_oloc_copy(&req->r_base_oloc, oloc);
4792 	req->r_flags = CEPH_OSD_FLAG_READ;
4793 
4794 	ret = ceph_osdc_alloc_messages(req, GFP_NOIO);
4795 	if (ret)
4796 		goto out_put_req;
4797 
4798 	pages = ceph_alloc_page_vector(1, GFP_NOIO);
4799 	if (IS_ERR(pages)) {
4800 		ret = PTR_ERR(pages);
4801 		goto out_put_req;
4802 	}
4803 
4804 	osd_req_op_init(req, 0, CEPH_OSD_OP_LIST_WATCHERS, 0);
4805 	ceph_osd_data_pages_init(osd_req_op_data(req, 0, list_watchers,
4806 						 response_data),
4807 				 pages, PAGE_SIZE, 0, false, true);
4808 
4809 	ceph_osdc_start_request(osdc, req, false);
4810 	ret = ceph_osdc_wait_request(osdc, req);
4811 	if (ret >= 0) {
4812 		void *p = page_address(pages[0]);
4813 		void *const end = p + req->r_ops[0].outdata_len;
4814 
4815 		ret = decode_watchers(&p, end, watchers, num_watchers);
4816 	}
4817 
4818 out_put_req:
4819 	ceph_osdc_put_request(req);
4820 	return ret;
4821 }
4822 EXPORT_SYMBOL(ceph_osdc_list_watchers);
4823 
4824 /*
4825  * Call all pending notify callbacks - for use after a watch is
4826  * unregistered, to make sure no more callbacks for it will be invoked
4827  */
4828 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
4829 {
4830 	dout("%s osdc %p\n", __func__, osdc);
4831 	flush_workqueue(osdc->notify_wq);
4832 }
4833 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
4834 
4835 void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc)
4836 {
4837 	down_read(&osdc->lock);
4838 	maybe_request_map(osdc);
4839 	up_read(&osdc->lock);
4840 }
4841 EXPORT_SYMBOL(ceph_osdc_maybe_request_map);
4842 
4843 /*
4844  * Execute an OSD class method on an object.
4845  *
4846  * @flags: CEPH_OSD_FLAG_*
4847  * @resp_len: in/out param for reply length
4848  */
4849 int ceph_osdc_call(struct ceph_osd_client *osdc,
4850 		   struct ceph_object_id *oid,
4851 		   struct ceph_object_locator *oloc,
4852 		   const char *class, const char *method,
4853 		   unsigned int flags,
4854 		   struct page *req_page, size_t req_len,
4855 		   struct page *resp_page, size_t *resp_len)
4856 {
4857 	struct ceph_osd_request *req;
4858 	int ret;
4859 
4860 	if (req_len > PAGE_SIZE || (resp_page && *resp_len > PAGE_SIZE))
4861 		return -E2BIG;
4862 
4863 	req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO);
4864 	if (!req)
4865 		return -ENOMEM;
4866 
4867 	ceph_oid_copy(&req->r_base_oid, oid);
4868 	ceph_oloc_copy(&req->r_base_oloc, oloc);
4869 	req->r_flags = flags;
4870 
4871 	ret = ceph_osdc_alloc_messages(req, GFP_NOIO);
4872 	if (ret)
4873 		goto out_put_req;
4874 
4875 	osd_req_op_cls_init(req, 0, CEPH_OSD_OP_CALL, class, method);
4876 	if (req_page)
4877 		osd_req_op_cls_request_data_pages(req, 0, &req_page, req_len,
4878 						  0, false, false);
4879 	if (resp_page)
4880 		osd_req_op_cls_response_data_pages(req, 0, &resp_page,
4881 						   *resp_len, 0, false, false);
4882 
4883 	ceph_osdc_start_request(osdc, req, false);
4884 	ret = ceph_osdc_wait_request(osdc, req);
4885 	if (ret >= 0) {
4886 		ret = req->r_ops[0].rval;
4887 		if (resp_page)
4888 			*resp_len = req->r_ops[0].outdata_len;
4889 	}
4890 
4891 out_put_req:
4892 	ceph_osdc_put_request(req);
4893 	return ret;
4894 }
4895 EXPORT_SYMBOL(ceph_osdc_call);
4896 
4897 /*
4898  * init, shutdown
4899  */
4900 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
4901 {
4902 	int err;
4903 
4904 	dout("init\n");
4905 	osdc->client = client;
4906 	init_rwsem(&osdc->lock);
4907 	osdc->osds = RB_ROOT;
4908 	INIT_LIST_HEAD(&osdc->osd_lru);
4909 	spin_lock_init(&osdc->osd_lru_lock);
4910 	osd_init(&osdc->homeless_osd);
4911 	osdc->homeless_osd.o_osdc = osdc;
4912 	osdc->homeless_osd.o_osd = CEPH_HOMELESS_OSD;
4913 	osdc->last_linger_id = CEPH_LINGER_ID_START;
4914 	osdc->linger_requests = RB_ROOT;
4915 	osdc->map_checks = RB_ROOT;
4916 	osdc->linger_map_checks = RB_ROOT;
4917 	INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
4918 	INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
4919 
4920 	err = -ENOMEM;
4921 	osdc->osdmap = ceph_osdmap_alloc();
4922 	if (!osdc->osdmap)
4923 		goto out;
4924 
4925 	osdc->req_mempool = mempool_create_slab_pool(10,
4926 						     ceph_osd_request_cache);
4927 	if (!osdc->req_mempool)
4928 		goto out_map;
4929 
4930 	err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
4931 				PAGE_SIZE, 10, true, "osd_op");
4932 	if (err < 0)
4933 		goto out_mempool;
4934 	err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
4935 				PAGE_SIZE, 10, true, "osd_op_reply");
4936 	if (err < 0)
4937 		goto out_msgpool;
4938 
4939 	err = -ENOMEM;
4940 	osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
4941 	if (!osdc->notify_wq)
4942 		goto out_msgpool_reply;
4943 
4944 	schedule_delayed_work(&osdc->timeout_work,
4945 			      osdc->client->options->osd_keepalive_timeout);
4946 	schedule_delayed_work(&osdc->osds_timeout_work,
4947 	    round_jiffies_relative(osdc->client->options->osd_idle_ttl));
4948 
4949 	return 0;
4950 
4951 out_msgpool_reply:
4952 	ceph_msgpool_destroy(&osdc->msgpool_op_reply);
4953 out_msgpool:
4954 	ceph_msgpool_destroy(&osdc->msgpool_op);
4955 out_mempool:
4956 	mempool_destroy(osdc->req_mempool);
4957 out_map:
4958 	ceph_osdmap_destroy(osdc->osdmap);
4959 out:
4960 	return err;
4961 }
4962 
4963 void ceph_osdc_stop(struct ceph_osd_client *osdc)
4964 {
4965 	flush_workqueue(osdc->notify_wq);
4966 	destroy_workqueue(osdc->notify_wq);
4967 	cancel_delayed_work_sync(&osdc->timeout_work);
4968 	cancel_delayed_work_sync(&osdc->osds_timeout_work);
4969 
4970 	down_write(&osdc->lock);
4971 	while (!RB_EMPTY_ROOT(&osdc->osds)) {
4972 		struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
4973 						struct ceph_osd, o_node);
4974 		close_osd(osd);
4975 	}
4976 	up_write(&osdc->lock);
4977 	WARN_ON(refcount_read(&osdc->homeless_osd.o_ref) != 1);
4978 	osd_cleanup(&osdc->homeless_osd);
4979 
4980 	WARN_ON(!list_empty(&osdc->osd_lru));
4981 	WARN_ON(!RB_EMPTY_ROOT(&osdc->linger_requests));
4982 	WARN_ON(!RB_EMPTY_ROOT(&osdc->map_checks));
4983 	WARN_ON(!RB_EMPTY_ROOT(&osdc->linger_map_checks));
4984 	WARN_ON(atomic_read(&osdc->num_requests));
4985 	WARN_ON(atomic_read(&osdc->num_homeless));
4986 
4987 	ceph_osdmap_destroy(osdc->osdmap);
4988 	mempool_destroy(osdc->req_mempool);
4989 	ceph_msgpool_destroy(&osdc->msgpool_op);
4990 	ceph_msgpool_destroy(&osdc->msgpool_op_reply);
4991 }
4992 
4993 /*
4994  * Read some contiguous pages.  If we cross a stripe boundary, shorten
4995  * *plen.  Return number of bytes read, or error.
4996  */
4997 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
4998 			struct ceph_vino vino, struct ceph_file_layout *layout,
4999 			u64 off, u64 *plen,
5000 			u32 truncate_seq, u64 truncate_size,
5001 			struct page **pages, int num_pages, int page_align)
5002 {
5003 	struct ceph_osd_request *req;
5004 	int rc = 0;
5005 
5006 	dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
5007 	     vino.snap, off, *plen);
5008 	req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1,
5009 				    CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
5010 				    NULL, truncate_seq, truncate_size,
5011 				    false);
5012 	if (IS_ERR(req))
5013 		return PTR_ERR(req);
5014 
5015 	/* it may be a short read due to an object boundary */
5016 	osd_req_op_extent_osd_data_pages(req, 0,
5017 				pages, *plen, page_align, false, false);
5018 
5019 	dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
5020 	     off, *plen, *plen, page_align);
5021 
5022 	rc = ceph_osdc_start_request(osdc, req, false);
5023 	if (!rc)
5024 		rc = ceph_osdc_wait_request(osdc, req);
5025 
5026 	ceph_osdc_put_request(req);
5027 	dout("readpages result %d\n", rc);
5028 	return rc;
5029 }
5030 EXPORT_SYMBOL(ceph_osdc_readpages);
5031 
5032 /*
5033  * do a synchronous write on N pages
5034  */
5035 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
5036 			 struct ceph_file_layout *layout,
5037 			 struct ceph_snap_context *snapc,
5038 			 u64 off, u64 len,
5039 			 u32 truncate_seq, u64 truncate_size,
5040 			 struct timespec *mtime,
5041 			 struct page **pages, int num_pages)
5042 {
5043 	struct ceph_osd_request *req;
5044 	int rc = 0;
5045 	int page_align = off & ~PAGE_MASK;
5046 
5047 	req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1,
5048 				    CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
5049 				    snapc, truncate_seq, truncate_size,
5050 				    true);
5051 	if (IS_ERR(req))
5052 		return PTR_ERR(req);
5053 
5054 	/* it may be a short write due to an object boundary */
5055 	osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
5056 				false, false);
5057 	dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
5058 
5059 	req->r_mtime = *mtime;
5060 	rc = ceph_osdc_start_request(osdc, req, true);
5061 	if (!rc)
5062 		rc = ceph_osdc_wait_request(osdc, req);
5063 
5064 	ceph_osdc_put_request(req);
5065 	if (rc == 0)
5066 		rc = len;
5067 	dout("writepages result %d\n", rc);
5068 	return rc;
5069 }
5070 EXPORT_SYMBOL(ceph_osdc_writepages);
5071 
5072 int ceph_osdc_setup(void)
5073 {
5074 	size_t size = sizeof(struct ceph_osd_request) +
5075 	    CEPH_OSD_SLAB_OPS * sizeof(struct ceph_osd_req_op);
5076 
5077 	BUG_ON(ceph_osd_request_cache);
5078 	ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", size,
5079 						   0, 0, NULL);
5080 
5081 	return ceph_osd_request_cache ? 0 : -ENOMEM;
5082 }
5083 EXPORT_SYMBOL(ceph_osdc_setup);
5084 
5085 void ceph_osdc_cleanup(void)
5086 {
5087 	BUG_ON(!ceph_osd_request_cache);
5088 	kmem_cache_destroy(ceph_osd_request_cache);
5089 	ceph_osd_request_cache = NULL;
5090 }
5091 EXPORT_SYMBOL(ceph_osdc_cleanup);
5092 
5093 /*
5094  * handle incoming message
5095  */
5096 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5097 {
5098 	struct ceph_osd *osd = con->private;
5099 	struct ceph_osd_client *osdc = osd->o_osdc;
5100 	int type = le16_to_cpu(msg->hdr.type);
5101 
5102 	switch (type) {
5103 	case CEPH_MSG_OSD_MAP:
5104 		ceph_osdc_handle_map(osdc, msg);
5105 		break;
5106 	case CEPH_MSG_OSD_OPREPLY:
5107 		handle_reply(osd, msg);
5108 		break;
5109 	case CEPH_MSG_OSD_BACKOFF:
5110 		handle_backoff(osd, msg);
5111 		break;
5112 	case CEPH_MSG_WATCH_NOTIFY:
5113 		handle_watch_notify(osdc, msg);
5114 		break;
5115 
5116 	default:
5117 		pr_err("received unknown message type %d %s\n", type,
5118 		       ceph_msg_type_name(type));
5119 	}
5120 
5121 	ceph_msg_put(msg);
5122 }
5123 
5124 /*
5125  * Lookup and return message for incoming reply.  Don't try to do
5126  * anything about a larger than preallocated data portion of the
5127  * message at the moment - for now, just skip the message.
5128  */
5129 static struct ceph_msg *get_reply(struct ceph_connection *con,
5130 				  struct ceph_msg_header *hdr,
5131 				  int *skip)
5132 {
5133 	struct ceph_osd *osd = con->private;
5134 	struct ceph_osd_client *osdc = osd->o_osdc;
5135 	struct ceph_msg *m = NULL;
5136 	struct ceph_osd_request *req;
5137 	int front_len = le32_to_cpu(hdr->front_len);
5138 	int data_len = le32_to_cpu(hdr->data_len);
5139 	u64 tid = le64_to_cpu(hdr->tid);
5140 
5141 	down_read(&osdc->lock);
5142 	if (!osd_registered(osd)) {
5143 		dout("%s osd%d unknown, skipping\n", __func__, osd->o_osd);
5144 		*skip = 1;
5145 		goto out_unlock_osdc;
5146 	}
5147 	WARN_ON(osd->o_osd != le64_to_cpu(hdr->src.num));
5148 
5149 	mutex_lock(&osd->lock);
5150 	req = lookup_request(&osd->o_requests, tid);
5151 	if (!req) {
5152 		dout("%s osd%d tid %llu unknown, skipping\n", __func__,
5153 		     osd->o_osd, tid);
5154 		*skip = 1;
5155 		goto out_unlock_session;
5156 	}
5157 
5158 	ceph_msg_revoke_incoming(req->r_reply);
5159 
5160 	if (front_len > req->r_reply->front_alloc_len) {
5161 		pr_warn("%s osd%d tid %llu front %d > preallocated %d\n",
5162 			__func__, osd->o_osd, req->r_tid, front_len,
5163 			req->r_reply->front_alloc_len);
5164 		m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
5165 				 false);
5166 		if (!m)
5167 			goto out_unlock_session;
5168 		ceph_msg_put(req->r_reply);
5169 		req->r_reply = m;
5170 	}
5171 
5172 	if (data_len > req->r_reply->data_length) {
5173 		pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n",
5174 			__func__, osd->o_osd, req->r_tid, data_len,
5175 			req->r_reply->data_length);
5176 		m = NULL;
5177 		*skip = 1;
5178 		goto out_unlock_session;
5179 	}
5180 
5181 	m = ceph_msg_get(req->r_reply);
5182 	dout("get_reply tid %lld %p\n", tid, m);
5183 
5184 out_unlock_session:
5185 	mutex_unlock(&osd->lock);
5186 out_unlock_osdc:
5187 	up_read(&osdc->lock);
5188 	return m;
5189 }
5190 
5191 /*
5192  * TODO: switch to a msg-owned pagelist
5193  */
5194 static struct ceph_msg *alloc_msg_with_page_vector(struct ceph_msg_header *hdr)
5195 {
5196 	struct ceph_msg *m;
5197 	int type = le16_to_cpu(hdr->type);
5198 	u32 front_len = le32_to_cpu(hdr->front_len);
5199 	u32 data_len = le32_to_cpu(hdr->data_len);
5200 
5201 	m = ceph_msg_new(type, front_len, GFP_NOIO, false);
5202 	if (!m)
5203 		return NULL;
5204 
5205 	if (data_len) {
5206 		struct page **pages;
5207 		struct ceph_osd_data osd_data;
5208 
5209 		pages = ceph_alloc_page_vector(calc_pages_for(0, data_len),
5210 					       GFP_NOIO);
5211 		if (IS_ERR(pages)) {
5212 			ceph_msg_put(m);
5213 			return NULL;
5214 		}
5215 
5216 		ceph_osd_data_pages_init(&osd_data, pages, data_len, 0, false,
5217 					 false);
5218 		ceph_osdc_msg_data_add(m, &osd_data);
5219 	}
5220 
5221 	return m;
5222 }
5223 
5224 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
5225 				  struct ceph_msg_header *hdr,
5226 				  int *skip)
5227 {
5228 	struct ceph_osd *osd = con->private;
5229 	int type = le16_to_cpu(hdr->type);
5230 
5231 	*skip = 0;
5232 	switch (type) {
5233 	case CEPH_MSG_OSD_MAP:
5234 	case CEPH_MSG_OSD_BACKOFF:
5235 	case CEPH_MSG_WATCH_NOTIFY:
5236 		return alloc_msg_with_page_vector(hdr);
5237 	case CEPH_MSG_OSD_OPREPLY:
5238 		return get_reply(con, hdr, skip);
5239 	default:
5240 		pr_warn("%s osd%d unknown msg type %d, skipping\n", __func__,
5241 			osd->o_osd, type);
5242 		*skip = 1;
5243 		return NULL;
5244 	}
5245 }
5246 
5247 /*
5248  * Wrappers to refcount containing ceph_osd struct
5249  */
5250 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
5251 {
5252 	struct ceph_osd *osd = con->private;
5253 	if (get_osd(osd))
5254 		return con;
5255 	return NULL;
5256 }
5257 
5258 static void put_osd_con(struct ceph_connection *con)
5259 {
5260 	struct ceph_osd *osd = con->private;
5261 	put_osd(osd);
5262 }
5263 
5264 /*
5265  * authentication
5266  */
5267 /*
5268  * Note: returned pointer is the address of a structure that's
5269  * managed separately.  Caller must *not* attempt to free it.
5270  */
5271 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
5272 					int *proto, int force_new)
5273 {
5274 	struct ceph_osd *o = con->private;
5275 	struct ceph_osd_client *osdc = o->o_osdc;
5276 	struct ceph_auth_client *ac = osdc->client->monc.auth;
5277 	struct ceph_auth_handshake *auth = &o->o_auth;
5278 
5279 	if (force_new && auth->authorizer) {
5280 		ceph_auth_destroy_authorizer(auth->authorizer);
5281 		auth->authorizer = NULL;
5282 	}
5283 	if (!auth->authorizer) {
5284 		int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
5285 						      auth);
5286 		if (ret)
5287 			return ERR_PTR(ret);
5288 	} else {
5289 		int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
5290 						     auth);
5291 		if (ret)
5292 			return ERR_PTR(ret);
5293 	}
5294 	*proto = ac->protocol;
5295 
5296 	return auth;
5297 }
5298 
5299 
5300 static int verify_authorizer_reply(struct ceph_connection *con)
5301 {
5302 	struct ceph_osd *o = con->private;
5303 	struct ceph_osd_client *osdc = o->o_osdc;
5304 	struct ceph_auth_client *ac = osdc->client->monc.auth;
5305 
5306 	return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer);
5307 }
5308 
5309 static int invalidate_authorizer(struct ceph_connection *con)
5310 {
5311 	struct ceph_osd *o = con->private;
5312 	struct ceph_osd_client *osdc = o->o_osdc;
5313 	struct ceph_auth_client *ac = osdc->client->monc.auth;
5314 
5315 	ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
5316 	return ceph_monc_validate_auth(&osdc->client->monc);
5317 }
5318 
5319 static void osd_reencode_message(struct ceph_msg *msg)
5320 {
5321 	int type = le16_to_cpu(msg->hdr.type);
5322 
5323 	if (type == CEPH_MSG_OSD_OP)
5324 		encode_request_finish(msg);
5325 }
5326 
5327 static int osd_sign_message(struct ceph_msg *msg)
5328 {
5329 	struct ceph_osd *o = msg->con->private;
5330 	struct ceph_auth_handshake *auth = &o->o_auth;
5331 
5332 	return ceph_auth_sign_message(auth, msg);
5333 }
5334 
5335 static int osd_check_message_signature(struct ceph_msg *msg)
5336 {
5337 	struct ceph_osd *o = msg->con->private;
5338 	struct ceph_auth_handshake *auth = &o->o_auth;
5339 
5340 	return ceph_auth_check_message_signature(auth, msg);
5341 }
5342 
5343 static const struct ceph_connection_operations osd_con_ops = {
5344 	.get = get_osd_con,
5345 	.put = put_osd_con,
5346 	.dispatch = dispatch,
5347 	.get_authorizer = get_authorizer,
5348 	.verify_authorizer_reply = verify_authorizer_reply,
5349 	.invalidate_authorizer = invalidate_authorizer,
5350 	.alloc_msg = alloc_msg,
5351 	.reencode_message = osd_reencode_message,
5352 	.sign_message = osd_sign_message,
5353 	.check_message_signature = osd_check_message_signature,
5354 	.fault = osd_fault,
5355 };
5356